LoopStrengthReduce.cpp revision 05fecbe42e835b30274a7b38af27687a8abbd114
1//===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This transformation analyzes and transforms the induction variables (and 11// computations derived from them) into forms suitable for efficient execution 12// on the target. 13// 14// This pass performs a strength reduction on array references inside loops that 15// have as one or more of their components the loop induction variable, it 16// rewrites expressions to take advantage of scaled-index addressing modes 17// available on the target, and it performs a variety of other optimizations 18// related to loop induction variables. 19// 20// Terminology note: this code has a lot of handling for "post-increment" or 21// "post-inc" users. This is not talking about post-increment addressing modes; 22// it is instead talking about code like this: 23// 24// %i = phi [ 0, %entry ], [ %i.next, %latch ] 25// ... 26// %i.next = add %i, 1 27// %c = icmp eq %i.next, %n 28// 29// The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however 30// it's useful to think about these as the same register, with some uses using 31// the value of the register before the add and some using // it after. In this 32// example, the icmp is a post-increment user, since it uses %i.next, which is 33// the value of the induction variable after the increment. The other common 34// case of post-increment users is users outside the loop. 35// 36// TODO: More sophistication in the way Formulae are generated and filtered. 37// 38// TODO: Handle multiple loops at a time. 39// 40// TODO: Should TargetLowering::AddrMode::BaseGV be changed to a ConstantExpr 41// instead of a GlobalValue? 42// 43// TODO: When truncation is free, truncate ICmp users' operands to make it a 44// smaller encoding (on x86 at least). 45// 46// TODO: When a negated register is used by an add (such as in a list of 47// multiple base registers, or as the increment expression in an addrec), 48// we may not actually need both reg and (-1 * reg) in registers; the 49// negation can be implemented by using a sub instead of an add. The 50// lack of support for taking this into consideration when making 51// register pressure decisions is partly worked around by the "Special" 52// use kind. 53// 54//===----------------------------------------------------------------------===// 55 56#define DEBUG_TYPE "loop-reduce" 57#include "llvm/Transforms/Scalar.h" 58#include "llvm/Constants.h" 59#include "llvm/Instructions.h" 60#include "llvm/IntrinsicInst.h" 61#include "llvm/DerivedTypes.h" 62#include "llvm/Analysis/IVUsers.h" 63#include "llvm/Analysis/Dominators.h" 64#include "llvm/Analysis/LoopPass.h" 65#include "llvm/Analysis/ScalarEvolutionExpander.h" 66#include "llvm/Assembly/Writer.h" 67#include "llvm/Transforms/Utils/BasicBlockUtils.h" 68#include "llvm/Transforms/Utils/Local.h" 69#include "llvm/ADT/SmallBitVector.h" 70#include "llvm/ADT/SetVector.h" 71#include "llvm/ADT/DenseSet.h" 72#include "llvm/Support/Debug.h" 73#include "llvm/Support/CommandLine.h" 74#include "llvm/Support/ValueHandle.h" 75#include "llvm/Support/raw_ostream.h" 76#include "llvm/Target/TargetLowering.h" 77#include <algorithm> 78using namespace llvm; 79 80// Temporary flag to cleanup congruent phis after LSR phi expansion. 81// It's currently disabled until we can determine whether it's truly useful or 82// not. The flag should be removed after the v3.0 release. 83// This is now needed for ivchains. 84static cl::opt<bool> EnablePhiElim( 85 "enable-lsr-phielim", cl::Hidden, cl::init(true), 86 cl::desc("Enable LSR phi elimination")); 87 88#ifndef NDEBUG 89// Stress test IV chain generation. 90static cl::opt<bool> StressIVChain( 91 "stress-ivchain", cl::Hidden, cl::init(false), 92 cl::desc("Stress test LSR IV chains")); 93#else 94static bool StressIVChain = false; 95#endif 96 97namespace { 98 99/// RegSortData - This class holds data which is used to order reuse candidates. 100class RegSortData { 101public: 102 /// UsedByIndices - This represents the set of LSRUse indices which reference 103 /// a particular register. 104 SmallBitVector UsedByIndices; 105 106 RegSortData() {} 107 108 void print(raw_ostream &OS) const; 109 void dump() const; 110}; 111 112} 113 114void RegSortData::print(raw_ostream &OS) const { 115 OS << "[NumUses=" << UsedByIndices.count() << ']'; 116} 117 118void RegSortData::dump() const { 119 print(errs()); errs() << '\n'; 120} 121 122namespace { 123 124/// RegUseTracker - Map register candidates to information about how they are 125/// used. 126class RegUseTracker { 127 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy; 128 129 RegUsesTy RegUsesMap; 130 SmallVector<const SCEV *, 16> RegSequence; 131 132public: 133 void CountRegister(const SCEV *Reg, size_t LUIdx); 134 void DropRegister(const SCEV *Reg, size_t LUIdx); 135 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx); 136 137 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const; 138 139 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const; 140 141 void clear(); 142 143 typedef SmallVectorImpl<const SCEV *>::iterator iterator; 144 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator; 145 iterator begin() { return RegSequence.begin(); } 146 iterator end() { return RegSequence.end(); } 147 const_iterator begin() const { return RegSequence.begin(); } 148 const_iterator end() const { return RegSequence.end(); } 149}; 150 151} 152 153void 154RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) { 155 std::pair<RegUsesTy::iterator, bool> Pair = 156 RegUsesMap.insert(std::make_pair(Reg, RegSortData())); 157 RegSortData &RSD = Pair.first->second; 158 if (Pair.second) 159 RegSequence.push_back(Reg); 160 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1)); 161 RSD.UsedByIndices.set(LUIdx); 162} 163 164void 165RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) { 166 RegUsesTy::iterator It = RegUsesMap.find(Reg); 167 assert(It != RegUsesMap.end()); 168 RegSortData &RSD = It->second; 169 assert(RSD.UsedByIndices.size() > LUIdx); 170 RSD.UsedByIndices.reset(LUIdx); 171} 172 173void 174RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) { 175 assert(LUIdx <= LastLUIdx); 176 177 // Update RegUses. The data structure is not optimized for this purpose; 178 // we must iterate through it and update each of the bit vectors. 179 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end(); 180 I != E; ++I) { 181 SmallBitVector &UsedByIndices = I->second.UsedByIndices; 182 if (LUIdx < UsedByIndices.size()) 183 UsedByIndices[LUIdx] = 184 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0; 185 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx)); 186 } 187} 188 189bool 190RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const { 191 RegUsesTy::const_iterator I = RegUsesMap.find(Reg); 192 if (I == RegUsesMap.end()) 193 return false; 194 const SmallBitVector &UsedByIndices = I->second.UsedByIndices; 195 int i = UsedByIndices.find_first(); 196 if (i == -1) return false; 197 if ((size_t)i != LUIdx) return true; 198 return UsedByIndices.find_next(i) != -1; 199} 200 201const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const { 202 RegUsesTy::const_iterator I = RegUsesMap.find(Reg); 203 assert(I != RegUsesMap.end() && "Unknown register!"); 204 return I->second.UsedByIndices; 205} 206 207void RegUseTracker::clear() { 208 RegUsesMap.clear(); 209 RegSequence.clear(); 210} 211 212namespace { 213 214/// Formula - This class holds information that describes a formula for 215/// computing satisfying a use. It may include broken-out immediates and scaled 216/// registers. 217struct Formula { 218 /// AM - This is used to represent complex addressing, as well as other kinds 219 /// of interesting uses. 220 TargetLowering::AddrMode AM; 221 222 /// BaseRegs - The list of "base" registers for this use. When this is 223 /// non-empty, AM.HasBaseReg should be set to true. 224 SmallVector<const SCEV *, 2> BaseRegs; 225 226 /// ScaledReg - The 'scaled' register for this use. This should be non-null 227 /// when AM.Scale is not zero. 228 const SCEV *ScaledReg; 229 230 /// UnfoldedOffset - An additional constant offset which added near the 231 /// use. This requires a temporary register, but the offset itself can 232 /// live in an add immediate field rather than a register. 233 int64_t UnfoldedOffset; 234 235 Formula() : ScaledReg(0), UnfoldedOffset(0) {} 236 237 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE); 238 239 unsigned getNumRegs() const; 240 Type *getType() const; 241 242 void DeleteBaseReg(const SCEV *&S); 243 244 bool referencesReg(const SCEV *S) const; 245 bool hasRegsUsedByUsesOtherThan(size_t LUIdx, 246 const RegUseTracker &RegUses) const; 247 248 void print(raw_ostream &OS) const; 249 void dump() const; 250}; 251 252} 253 254/// DoInitialMatch - Recursion helper for InitialMatch. 255static void DoInitialMatch(const SCEV *S, Loop *L, 256 SmallVectorImpl<const SCEV *> &Good, 257 SmallVectorImpl<const SCEV *> &Bad, 258 ScalarEvolution &SE) { 259 // Collect expressions which properly dominate the loop header. 260 if (SE.properlyDominates(S, L->getHeader())) { 261 Good.push_back(S); 262 return; 263 } 264 265 // Look at add operands. 266 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 267 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 268 I != E; ++I) 269 DoInitialMatch(*I, L, Good, Bad, SE); 270 return; 271 } 272 273 // Look at addrec operands. 274 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) 275 if (!AR->getStart()->isZero()) { 276 DoInitialMatch(AR->getStart(), L, Good, Bad, SE); 277 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0), 278 AR->getStepRecurrence(SE), 279 // FIXME: AR->getNoWrapFlags() 280 AR->getLoop(), SCEV::FlagAnyWrap), 281 L, Good, Bad, SE); 282 return; 283 } 284 285 // Handle a multiplication by -1 (negation) if it didn't fold. 286 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) 287 if (Mul->getOperand(0)->isAllOnesValue()) { 288 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end()); 289 const SCEV *NewMul = SE.getMulExpr(Ops); 290 291 SmallVector<const SCEV *, 4> MyGood; 292 SmallVector<const SCEV *, 4> MyBad; 293 DoInitialMatch(NewMul, L, MyGood, MyBad, SE); 294 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue( 295 SE.getEffectiveSCEVType(NewMul->getType()))); 296 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(), 297 E = MyGood.end(); I != E; ++I) 298 Good.push_back(SE.getMulExpr(NegOne, *I)); 299 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(), 300 E = MyBad.end(); I != E; ++I) 301 Bad.push_back(SE.getMulExpr(NegOne, *I)); 302 return; 303 } 304 305 // Ok, we can't do anything interesting. Just stuff the whole thing into a 306 // register and hope for the best. 307 Bad.push_back(S); 308} 309 310/// InitialMatch - Incorporate loop-variant parts of S into this Formula, 311/// attempting to keep all loop-invariant and loop-computable values in a 312/// single base register. 313void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) { 314 SmallVector<const SCEV *, 4> Good; 315 SmallVector<const SCEV *, 4> Bad; 316 DoInitialMatch(S, L, Good, Bad, SE); 317 if (!Good.empty()) { 318 const SCEV *Sum = SE.getAddExpr(Good); 319 if (!Sum->isZero()) 320 BaseRegs.push_back(Sum); 321 AM.HasBaseReg = true; 322 } 323 if (!Bad.empty()) { 324 const SCEV *Sum = SE.getAddExpr(Bad); 325 if (!Sum->isZero()) 326 BaseRegs.push_back(Sum); 327 AM.HasBaseReg = true; 328 } 329} 330 331/// getNumRegs - Return the total number of register operands used by this 332/// formula. This does not include register uses implied by non-constant 333/// addrec strides. 334unsigned Formula::getNumRegs() const { 335 return !!ScaledReg + BaseRegs.size(); 336} 337 338/// getType - Return the type of this formula, if it has one, or null 339/// otherwise. This type is meaningless except for the bit size. 340Type *Formula::getType() const { 341 return !BaseRegs.empty() ? BaseRegs.front()->getType() : 342 ScaledReg ? ScaledReg->getType() : 343 AM.BaseGV ? AM.BaseGV->getType() : 344 0; 345} 346 347/// DeleteBaseReg - Delete the given base reg from the BaseRegs list. 348void Formula::DeleteBaseReg(const SCEV *&S) { 349 if (&S != &BaseRegs.back()) 350 std::swap(S, BaseRegs.back()); 351 BaseRegs.pop_back(); 352} 353 354/// referencesReg - Test if this formula references the given register. 355bool Formula::referencesReg(const SCEV *S) const { 356 return S == ScaledReg || 357 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end(); 358} 359 360/// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers 361/// which are used by uses other than the use with the given index. 362bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx, 363 const RegUseTracker &RegUses) const { 364 if (ScaledReg) 365 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx)) 366 return true; 367 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(), 368 E = BaseRegs.end(); I != E; ++I) 369 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx)) 370 return true; 371 return false; 372} 373 374void Formula::print(raw_ostream &OS) const { 375 bool First = true; 376 if (AM.BaseGV) { 377 if (!First) OS << " + "; else First = false; 378 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false); 379 } 380 if (AM.BaseOffs != 0) { 381 if (!First) OS << " + "; else First = false; 382 OS << AM.BaseOffs; 383 } 384 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(), 385 E = BaseRegs.end(); I != E; ++I) { 386 if (!First) OS << " + "; else First = false; 387 OS << "reg(" << **I << ')'; 388 } 389 if (AM.HasBaseReg && BaseRegs.empty()) { 390 if (!First) OS << " + "; else First = false; 391 OS << "**error: HasBaseReg**"; 392 } else if (!AM.HasBaseReg && !BaseRegs.empty()) { 393 if (!First) OS << " + "; else First = false; 394 OS << "**error: !HasBaseReg**"; 395 } 396 if (AM.Scale != 0) { 397 if (!First) OS << " + "; else First = false; 398 OS << AM.Scale << "*reg("; 399 if (ScaledReg) 400 OS << *ScaledReg; 401 else 402 OS << "<unknown>"; 403 OS << ')'; 404 } 405 if (UnfoldedOffset != 0) { 406 if (!First) OS << " + "; else First = false; 407 OS << "imm(" << UnfoldedOffset << ')'; 408 } 409} 410 411void Formula::dump() const { 412 print(errs()); errs() << '\n'; 413} 414 415/// isAddRecSExtable - Return true if the given addrec can be sign-extended 416/// without changing its value. 417static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) { 418 Type *WideTy = 419 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1); 420 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy)); 421} 422 423/// isAddSExtable - Return true if the given add can be sign-extended 424/// without changing its value. 425static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) { 426 Type *WideTy = 427 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1); 428 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy)); 429} 430 431/// isMulSExtable - Return true if the given mul can be sign-extended 432/// without changing its value. 433static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) { 434 Type *WideTy = 435 IntegerType::get(SE.getContext(), 436 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands()); 437 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy)); 438} 439 440/// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined 441/// and if the remainder is known to be zero, or null otherwise. If 442/// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified 443/// to Y, ignoring that the multiplication may overflow, which is useful when 444/// the result will be used in a context where the most significant bits are 445/// ignored. 446static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS, 447 ScalarEvolution &SE, 448 bool IgnoreSignificantBits = false) { 449 // Handle the trivial case, which works for any SCEV type. 450 if (LHS == RHS) 451 return SE.getConstant(LHS->getType(), 1); 452 453 // Handle a few RHS special cases. 454 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS); 455 if (RC) { 456 const APInt &RA = RC->getValue()->getValue(); 457 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do 458 // some folding. 459 if (RA.isAllOnesValue()) 460 return SE.getMulExpr(LHS, RC); 461 // Handle x /s 1 as x. 462 if (RA == 1) 463 return LHS; 464 } 465 466 // Check for a division of a constant by a constant. 467 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) { 468 if (!RC) 469 return 0; 470 const APInt &LA = C->getValue()->getValue(); 471 const APInt &RA = RC->getValue()->getValue(); 472 if (LA.srem(RA) != 0) 473 return 0; 474 return SE.getConstant(LA.sdiv(RA)); 475 } 476 477 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow. 478 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) { 479 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) { 480 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE, 481 IgnoreSignificantBits); 482 if (!Step) return 0; 483 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE, 484 IgnoreSignificantBits); 485 if (!Start) return 0; 486 // FlagNW is independent of the start value, step direction, and is 487 // preserved with smaller magnitude steps. 488 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 489 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap); 490 } 491 return 0; 492 } 493 494 // Distribute the sdiv over add operands, if the add doesn't overflow. 495 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) { 496 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) { 497 SmallVector<const SCEV *, 8> Ops; 498 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 499 I != E; ++I) { 500 const SCEV *Op = getExactSDiv(*I, RHS, SE, 501 IgnoreSignificantBits); 502 if (!Op) return 0; 503 Ops.push_back(Op); 504 } 505 return SE.getAddExpr(Ops); 506 } 507 return 0; 508 } 509 510 // Check for a multiply operand that we can pull RHS out of. 511 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) { 512 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) { 513 SmallVector<const SCEV *, 4> Ops; 514 bool Found = false; 515 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end(); 516 I != E; ++I) { 517 const SCEV *S = *I; 518 if (!Found) 519 if (const SCEV *Q = getExactSDiv(S, RHS, SE, 520 IgnoreSignificantBits)) { 521 S = Q; 522 Found = true; 523 } 524 Ops.push_back(S); 525 } 526 return Found ? SE.getMulExpr(Ops) : 0; 527 } 528 return 0; 529 } 530 531 // Otherwise we don't know. 532 return 0; 533} 534 535/// ExtractImmediate - If S involves the addition of a constant integer value, 536/// return that integer value, and mutate S to point to a new SCEV with that 537/// value excluded. 538static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) { 539 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) { 540 if (C->getValue()->getValue().getMinSignedBits() <= 64) { 541 S = SE.getConstant(C->getType(), 0); 542 return C->getValue()->getSExtValue(); 543 } 544 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 545 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end()); 546 int64_t Result = ExtractImmediate(NewOps.front(), SE); 547 if (Result != 0) 548 S = SE.getAddExpr(NewOps); 549 return Result; 550 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 551 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end()); 552 int64_t Result = ExtractImmediate(NewOps.front(), SE); 553 if (Result != 0) 554 S = SE.getAddRecExpr(NewOps, AR->getLoop(), 555 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 556 SCEV::FlagAnyWrap); 557 return Result; 558 } 559 return 0; 560} 561 562/// ExtractSymbol - If S involves the addition of a GlobalValue address, 563/// return that symbol, and mutate S to point to a new SCEV with that 564/// value excluded. 565static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) { 566 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 567 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) { 568 S = SE.getConstant(GV->getType(), 0); 569 return GV; 570 } 571 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 572 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end()); 573 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE); 574 if (Result) 575 S = SE.getAddExpr(NewOps); 576 return Result; 577 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 578 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end()); 579 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE); 580 if (Result) 581 S = SE.getAddRecExpr(NewOps, AR->getLoop(), 582 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 583 SCEV::FlagAnyWrap); 584 return Result; 585 } 586 return 0; 587} 588 589/// isAddressUse - Returns true if the specified instruction is using the 590/// specified value as an address. 591static bool isAddressUse(Instruction *Inst, Value *OperandVal) { 592 bool isAddress = isa<LoadInst>(Inst); 593 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 594 if (SI->getOperand(1) == OperandVal) 595 isAddress = true; 596 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 597 // Addressing modes can also be folded into prefetches and a variety 598 // of intrinsics. 599 switch (II->getIntrinsicID()) { 600 default: break; 601 case Intrinsic::prefetch: 602 case Intrinsic::x86_sse_storeu_ps: 603 case Intrinsic::x86_sse2_storeu_pd: 604 case Intrinsic::x86_sse2_storeu_dq: 605 case Intrinsic::x86_sse2_storel_dq: 606 if (II->getArgOperand(0) == OperandVal) 607 isAddress = true; 608 break; 609 } 610 } 611 return isAddress; 612} 613 614/// getAccessType - Return the type of the memory being accessed. 615static Type *getAccessType(const Instruction *Inst) { 616 Type *AccessTy = Inst->getType(); 617 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) 618 AccessTy = SI->getOperand(0)->getType(); 619 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 620 // Addressing modes can also be folded into prefetches and a variety 621 // of intrinsics. 622 switch (II->getIntrinsicID()) { 623 default: break; 624 case Intrinsic::x86_sse_storeu_ps: 625 case Intrinsic::x86_sse2_storeu_pd: 626 case Intrinsic::x86_sse2_storeu_dq: 627 case Intrinsic::x86_sse2_storel_dq: 628 AccessTy = II->getArgOperand(0)->getType(); 629 break; 630 } 631 } 632 633 // All pointers have the same requirements, so canonicalize them to an 634 // arbitrary pointer type to minimize variation. 635 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy)) 636 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1), 637 PTy->getAddressSpace()); 638 639 return AccessTy; 640} 641 642/// isExistingPhi - Return true if this AddRec is already a phi in its loop. 643static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) { 644 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin(); 645 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 646 if (SE.isSCEVable(PN->getType()) && 647 (SE.getEffectiveSCEVType(PN->getType()) == 648 SE.getEffectiveSCEVType(AR->getType())) && 649 SE.getSCEV(PN) == AR) 650 return true; 651 } 652 return false; 653} 654 655/// Check if expanding this expression is likely to incur significant cost. This 656/// is tricky because SCEV doesn't track which expressions are actually computed 657/// by the current IR. 658/// 659/// We currently allow expansion of IV increments that involve adds, 660/// multiplication by constants, and AddRecs from existing phis. 661/// 662/// TODO: Allow UDivExpr if we can find an existing IV increment that is an 663/// obvious multiple of the UDivExpr. 664static bool isHighCostExpansion(const SCEV *S, 665 SmallPtrSet<const SCEV*, 8> &Processed, 666 ScalarEvolution &SE) { 667 // Zero/One operand expressions 668 switch (S->getSCEVType()) { 669 case scUnknown: 670 case scConstant: 671 return false; 672 case scTruncate: 673 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(), 674 Processed, SE); 675 case scZeroExtend: 676 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(), 677 Processed, SE); 678 case scSignExtend: 679 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(), 680 Processed, SE); 681 } 682 683 if (!Processed.insert(S)) 684 return false; 685 686 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 687 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 688 I != E; ++I) { 689 if (isHighCostExpansion(*I, Processed, SE)) 690 return true; 691 } 692 return false; 693 } 694 695 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) { 696 if (Mul->getNumOperands() == 2) { 697 // Multiplication by a constant is ok 698 if (isa<SCEVConstant>(Mul->getOperand(0))) 699 return isHighCostExpansion(Mul->getOperand(1), Processed, SE); 700 701 // If we have the value of one operand, check if an existing 702 // multiplication already generates this expression. 703 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) { 704 Value *UVal = U->getValue(); 705 for (Value::use_iterator UI = UVal->use_begin(), UE = UVal->use_end(); 706 UI != UE; ++UI) { 707 // If U is a constant, it may be used by a ConstantExpr. 708 Instruction *User = dyn_cast<Instruction>(*UI); 709 if (User && User->getOpcode() == Instruction::Mul 710 && SE.isSCEVable(User->getType())) { 711 return SE.getSCEV(User) == Mul; 712 } 713 } 714 } 715 } 716 } 717 718 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 719 if (isExistingPhi(AR, SE)) 720 return false; 721 } 722 723 // Fow now, consider any other type of expression (div/mul/min/max) high cost. 724 return true; 725} 726 727/// DeleteTriviallyDeadInstructions - If any of the instructions is the 728/// specified set are trivially dead, delete them and see if this makes any of 729/// their operands subsequently dead. 730static bool 731DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) { 732 bool Changed = false; 733 734 while (!DeadInsts.empty()) { 735 Instruction *I = dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()); 736 737 if (I == 0 || !isInstructionTriviallyDead(I)) 738 continue; 739 740 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 741 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 742 *OI = 0; 743 if (U->use_empty()) 744 DeadInsts.push_back(U); 745 } 746 747 I->eraseFromParent(); 748 Changed = true; 749 } 750 751 return Changed; 752} 753 754namespace { 755 756/// Cost - This class is used to measure and compare candidate formulae. 757class Cost { 758 /// TODO: Some of these could be merged. Also, a lexical ordering 759 /// isn't always optimal. 760 unsigned NumRegs; 761 unsigned AddRecCost; 762 unsigned NumIVMuls; 763 unsigned NumBaseAdds; 764 unsigned ImmCost; 765 unsigned SetupCost; 766 767public: 768 Cost() 769 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0), 770 SetupCost(0) {} 771 772 bool operator<(const Cost &Other) const; 773 774 void Loose(); 775 776#ifndef NDEBUG 777 // Once any of the metrics loses, they must all remain losers. 778 bool isValid() { 779 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds 780 | ImmCost | SetupCost) != ~0u) 781 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds 782 & ImmCost & SetupCost) == ~0u); 783 } 784#endif 785 786 bool isLoser() { 787 assert(isValid() && "invalid cost"); 788 return NumRegs == ~0u; 789 } 790 791 void RateFormula(const Formula &F, 792 SmallPtrSet<const SCEV *, 16> &Regs, 793 const DenseSet<const SCEV *> &VisitedRegs, 794 const Loop *L, 795 const SmallVectorImpl<int64_t> &Offsets, 796 ScalarEvolution &SE, DominatorTree &DT, 797 SmallPtrSet<const SCEV *, 16> *LoserRegs = 0); 798 799 void print(raw_ostream &OS) const; 800 void dump() const; 801 802private: 803 void RateRegister(const SCEV *Reg, 804 SmallPtrSet<const SCEV *, 16> &Regs, 805 const Loop *L, 806 ScalarEvolution &SE, DominatorTree &DT); 807 void RatePrimaryRegister(const SCEV *Reg, 808 SmallPtrSet<const SCEV *, 16> &Regs, 809 const Loop *L, 810 ScalarEvolution &SE, DominatorTree &DT, 811 SmallPtrSet<const SCEV *, 16> *LoserRegs); 812}; 813 814} 815 816/// RateRegister - Tally up interesting quantities from the given register. 817void Cost::RateRegister(const SCEV *Reg, 818 SmallPtrSet<const SCEV *, 16> &Regs, 819 const Loop *L, 820 ScalarEvolution &SE, DominatorTree &DT) { 821 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) { 822 // If this is an addrec for another loop, don't second-guess its addrec phi 823 // nodes. LSR isn't currently smart enough to reason about more than one 824 // loop at a time. LSR has already run on inner loops, will not run on outer 825 // loops, and cannot be expected to change sibling loops. 826 if (AR->getLoop() != L) { 827 // If the AddRec exists, consider it's register free and leave it alone. 828 if (isExistingPhi(AR, SE)) 829 return; 830 831 // Otherwise, do not consider this formula at all. 832 Loose(); 833 return; 834 } 835 AddRecCost += 1; /// TODO: This should be a function of the stride. 836 837 // Add the step value register, if it needs one. 838 // TODO: The non-affine case isn't precisely modeled here. 839 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) { 840 if (!Regs.count(AR->getOperand(1))) { 841 RateRegister(AR->getOperand(1), Regs, L, SE, DT); 842 if (isLoser()) 843 return; 844 } 845 } 846 } 847 ++NumRegs; 848 849 // Rough heuristic; favor registers which don't require extra setup 850 // instructions in the preheader. 851 if (!isa<SCEVUnknown>(Reg) && 852 !isa<SCEVConstant>(Reg) && 853 !(isa<SCEVAddRecExpr>(Reg) && 854 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) || 855 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart())))) 856 ++SetupCost; 857 858 NumIVMuls += isa<SCEVMulExpr>(Reg) && 859 SE.hasComputableLoopEvolution(Reg, L); 860} 861 862/// RatePrimaryRegister - Record this register in the set. If we haven't seen it 863/// before, rate it. Optional LoserRegs provides a way to declare any formula 864/// that refers to one of those regs an instant loser. 865void Cost::RatePrimaryRegister(const SCEV *Reg, 866 SmallPtrSet<const SCEV *, 16> &Regs, 867 const Loop *L, 868 ScalarEvolution &SE, DominatorTree &DT, 869 SmallPtrSet<const SCEV *, 16> *LoserRegs) { 870 if (LoserRegs && LoserRegs->count(Reg)) { 871 Loose(); 872 return; 873 } 874 if (Regs.insert(Reg)) { 875 RateRegister(Reg, Regs, L, SE, DT); 876 if (isLoser()) 877 LoserRegs->insert(Reg); 878 } 879} 880 881void Cost::RateFormula(const Formula &F, 882 SmallPtrSet<const SCEV *, 16> &Regs, 883 const DenseSet<const SCEV *> &VisitedRegs, 884 const Loop *L, 885 const SmallVectorImpl<int64_t> &Offsets, 886 ScalarEvolution &SE, DominatorTree &DT, 887 SmallPtrSet<const SCEV *, 16> *LoserRegs) { 888 // Tally up the registers. 889 if (const SCEV *ScaledReg = F.ScaledReg) { 890 if (VisitedRegs.count(ScaledReg)) { 891 Loose(); 892 return; 893 } 894 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs); 895 if (isLoser()) 896 return; 897 } 898 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(), 899 E = F.BaseRegs.end(); I != E; ++I) { 900 const SCEV *BaseReg = *I; 901 if (VisitedRegs.count(BaseReg)) { 902 Loose(); 903 return; 904 } 905 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs); 906 if (isLoser()) 907 return; 908 } 909 910 // Determine how many (unfolded) adds we'll need inside the loop. 911 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0); 912 if (NumBaseParts > 1) 913 NumBaseAdds += NumBaseParts - 1; 914 915 // Tally up the non-zero immediates. 916 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(), 917 E = Offsets.end(); I != E; ++I) { 918 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs; 919 if (F.AM.BaseGV) 920 ImmCost += 64; // Handle symbolic values conservatively. 921 // TODO: This should probably be the pointer size. 922 else if (Offset != 0) 923 ImmCost += APInt(64, Offset, true).getMinSignedBits(); 924 } 925 assert(isValid() && "invalid cost"); 926} 927 928/// Loose - Set this cost to a losing value. 929void Cost::Loose() { 930 NumRegs = ~0u; 931 AddRecCost = ~0u; 932 NumIVMuls = ~0u; 933 NumBaseAdds = ~0u; 934 ImmCost = ~0u; 935 SetupCost = ~0u; 936} 937 938/// operator< - Choose the lower cost. 939bool Cost::operator<(const Cost &Other) const { 940 if (NumRegs != Other.NumRegs) 941 return NumRegs < Other.NumRegs; 942 if (AddRecCost != Other.AddRecCost) 943 return AddRecCost < Other.AddRecCost; 944 if (NumIVMuls != Other.NumIVMuls) 945 return NumIVMuls < Other.NumIVMuls; 946 if (NumBaseAdds != Other.NumBaseAdds) 947 return NumBaseAdds < Other.NumBaseAdds; 948 if (ImmCost != Other.ImmCost) 949 return ImmCost < Other.ImmCost; 950 if (SetupCost != Other.SetupCost) 951 return SetupCost < Other.SetupCost; 952 return false; 953} 954 955void Cost::print(raw_ostream &OS) const { 956 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s"); 957 if (AddRecCost != 0) 958 OS << ", with addrec cost " << AddRecCost; 959 if (NumIVMuls != 0) 960 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s"); 961 if (NumBaseAdds != 0) 962 OS << ", plus " << NumBaseAdds << " base add" 963 << (NumBaseAdds == 1 ? "" : "s"); 964 if (ImmCost != 0) 965 OS << ", plus " << ImmCost << " imm cost"; 966 if (SetupCost != 0) 967 OS << ", plus " << SetupCost << " setup cost"; 968} 969 970void Cost::dump() const { 971 print(errs()); errs() << '\n'; 972} 973 974namespace { 975 976/// LSRFixup - An operand value in an instruction which is to be replaced 977/// with some equivalent, possibly strength-reduced, replacement. 978struct LSRFixup { 979 /// UserInst - The instruction which will be updated. 980 Instruction *UserInst; 981 982 /// OperandValToReplace - The operand of the instruction which will 983 /// be replaced. The operand may be used more than once; every instance 984 /// will be replaced. 985 Value *OperandValToReplace; 986 987 /// PostIncLoops - If this user is to use the post-incremented value of an 988 /// induction variable, this variable is non-null and holds the loop 989 /// associated with the induction variable. 990 PostIncLoopSet PostIncLoops; 991 992 /// LUIdx - The index of the LSRUse describing the expression which 993 /// this fixup needs, minus an offset (below). 994 size_t LUIdx; 995 996 /// Offset - A constant offset to be added to the LSRUse expression. 997 /// This allows multiple fixups to share the same LSRUse with different 998 /// offsets, for example in an unrolled loop. 999 int64_t Offset; 1000 1001 bool isUseFullyOutsideLoop(const Loop *L) const; 1002 1003 LSRFixup(); 1004 1005 void print(raw_ostream &OS) const; 1006 void dump() const; 1007}; 1008 1009} 1010 1011LSRFixup::LSRFixup() 1012 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {} 1013 1014/// isUseFullyOutsideLoop - Test whether this fixup always uses its 1015/// value outside of the given loop. 1016bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const { 1017 // PHI nodes use their value in their incoming blocks. 1018 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) { 1019 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 1020 if (PN->getIncomingValue(i) == OperandValToReplace && 1021 L->contains(PN->getIncomingBlock(i))) 1022 return false; 1023 return true; 1024 } 1025 1026 return !L->contains(UserInst); 1027} 1028 1029void LSRFixup::print(raw_ostream &OS) const { 1030 OS << "UserInst="; 1031 // Store is common and interesting enough to be worth special-casing. 1032 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) { 1033 OS << "store "; 1034 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false); 1035 } else if (UserInst->getType()->isVoidTy()) 1036 OS << UserInst->getOpcodeName(); 1037 else 1038 WriteAsOperand(OS, UserInst, /*PrintType=*/false); 1039 1040 OS << ", OperandValToReplace="; 1041 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false); 1042 1043 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(), 1044 E = PostIncLoops.end(); I != E; ++I) { 1045 OS << ", PostIncLoop="; 1046 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false); 1047 } 1048 1049 if (LUIdx != ~size_t(0)) 1050 OS << ", LUIdx=" << LUIdx; 1051 1052 if (Offset != 0) 1053 OS << ", Offset=" << Offset; 1054} 1055 1056void LSRFixup::dump() const { 1057 print(errs()); errs() << '\n'; 1058} 1059 1060namespace { 1061 1062/// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding 1063/// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*. 1064struct UniquifierDenseMapInfo { 1065 static SmallVector<const SCEV *, 2> getEmptyKey() { 1066 SmallVector<const SCEV *, 2> V; 1067 V.push_back(reinterpret_cast<const SCEV *>(-1)); 1068 return V; 1069 } 1070 1071 static SmallVector<const SCEV *, 2> getTombstoneKey() { 1072 SmallVector<const SCEV *, 2> V; 1073 V.push_back(reinterpret_cast<const SCEV *>(-2)); 1074 return V; 1075 } 1076 1077 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) { 1078 unsigned Result = 0; 1079 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(), 1080 E = V.end(); I != E; ++I) 1081 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I); 1082 return Result; 1083 } 1084 1085 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS, 1086 const SmallVector<const SCEV *, 2> &RHS) { 1087 return LHS == RHS; 1088 } 1089}; 1090 1091/// LSRUse - This class holds the state that LSR keeps for each use in 1092/// IVUsers, as well as uses invented by LSR itself. It includes information 1093/// about what kinds of things can be folded into the user, information about 1094/// the user itself, and information about how the use may be satisfied. 1095/// TODO: Represent multiple users of the same expression in common? 1096class LSRUse { 1097 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier; 1098 1099public: 1100 /// KindType - An enum for a kind of use, indicating what types of 1101 /// scaled and immediate operands it might support. 1102 enum KindType { 1103 Basic, ///< A normal use, with no folding. 1104 Special, ///< A special case of basic, allowing -1 scales. 1105 Address, ///< An address use; folding according to TargetLowering 1106 ICmpZero ///< An equality icmp with both operands folded into one. 1107 // TODO: Add a generic icmp too? 1108 }; 1109 1110 KindType Kind; 1111 Type *AccessTy; 1112 1113 SmallVector<int64_t, 8> Offsets; 1114 int64_t MinOffset; 1115 int64_t MaxOffset; 1116 1117 /// AllFixupsOutsideLoop - This records whether all of the fixups using this 1118 /// LSRUse are outside of the loop, in which case some special-case heuristics 1119 /// may be used. 1120 bool AllFixupsOutsideLoop; 1121 1122 /// WidestFixupType - This records the widest use type for any fixup using 1123 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different 1124 /// max fixup widths to be equivalent, because the narrower one may be relying 1125 /// on the implicit truncation to truncate away bogus bits. 1126 Type *WidestFixupType; 1127 1128 /// Formulae - A list of ways to build a value that can satisfy this user. 1129 /// After the list is populated, one of these is selected heuristically and 1130 /// used to formulate a replacement for OperandValToReplace in UserInst. 1131 SmallVector<Formula, 12> Formulae; 1132 1133 /// Regs - The set of register candidates used by all formulae in this LSRUse. 1134 SmallPtrSet<const SCEV *, 4> Regs; 1135 1136 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T), 1137 MinOffset(INT64_MAX), 1138 MaxOffset(INT64_MIN), 1139 AllFixupsOutsideLoop(true), 1140 WidestFixupType(0) {} 1141 1142 bool HasFormulaWithSameRegs(const Formula &F) const; 1143 bool InsertFormula(const Formula &F); 1144 void DeleteFormula(Formula &F); 1145 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses); 1146 1147 void print(raw_ostream &OS) const; 1148 void dump() const; 1149}; 1150 1151} 1152 1153/// HasFormula - Test whether this use as a formula which has the same 1154/// registers as the given formula. 1155bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const { 1156 SmallVector<const SCEV *, 2> Key = F.BaseRegs; 1157 if (F.ScaledReg) Key.push_back(F.ScaledReg); 1158 // Unstable sort by host order ok, because this is only used for uniquifying. 1159 std::sort(Key.begin(), Key.end()); 1160 return Uniquifier.count(Key); 1161} 1162 1163/// InsertFormula - If the given formula has not yet been inserted, add it to 1164/// the list, and return true. Return false otherwise. 1165bool LSRUse::InsertFormula(const Formula &F) { 1166 SmallVector<const SCEV *, 2> Key = F.BaseRegs; 1167 if (F.ScaledReg) Key.push_back(F.ScaledReg); 1168 // Unstable sort by host order ok, because this is only used for uniquifying. 1169 std::sort(Key.begin(), Key.end()); 1170 1171 if (!Uniquifier.insert(Key).second) 1172 return false; 1173 1174 // Using a register to hold the value of 0 is not profitable. 1175 assert((!F.ScaledReg || !F.ScaledReg->isZero()) && 1176 "Zero allocated in a scaled register!"); 1177#ifndef NDEBUG 1178 for (SmallVectorImpl<const SCEV *>::const_iterator I = 1179 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) 1180 assert(!(*I)->isZero() && "Zero allocated in a base register!"); 1181#endif 1182 1183 // Add the formula to the list. 1184 Formulae.push_back(F); 1185 1186 // Record registers now being used by this use. 1187 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end()); 1188 1189 return true; 1190} 1191 1192/// DeleteFormula - Remove the given formula from this use's list. 1193void LSRUse::DeleteFormula(Formula &F) { 1194 if (&F != &Formulae.back()) 1195 std::swap(F, Formulae.back()); 1196 Formulae.pop_back(); 1197} 1198 1199/// RecomputeRegs - Recompute the Regs field, and update RegUses. 1200void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) { 1201 // Now that we've filtered out some formulae, recompute the Regs set. 1202 SmallPtrSet<const SCEV *, 4> OldRegs = Regs; 1203 Regs.clear(); 1204 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(), 1205 E = Formulae.end(); I != E; ++I) { 1206 const Formula &F = *I; 1207 if (F.ScaledReg) Regs.insert(F.ScaledReg); 1208 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end()); 1209 } 1210 1211 // Update the RegTracker. 1212 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(), 1213 E = OldRegs.end(); I != E; ++I) 1214 if (!Regs.count(*I)) 1215 RegUses.DropRegister(*I, LUIdx); 1216} 1217 1218void LSRUse::print(raw_ostream &OS) const { 1219 OS << "LSR Use: Kind="; 1220 switch (Kind) { 1221 case Basic: OS << "Basic"; break; 1222 case Special: OS << "Special"; break; 1223 case ICmpZero: OS << "ICmpZero"; break; 1224 case Address: 1225 OS << "Address of "; 1226 if (AccessTy->isPointerTy()) 1227 OS << "pointer"; // the full pointer type could be really verbose 1228 else 1229 OS << *AccessTy; 1230 } 1231 1232 OS << ", Offsets={"; 1233 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(), 1234 E = Offsets.end(); I != E; ++I) { 1235 OS << *I; 1236 if (llvm::next(I) != E) 1237 OS << ','; 1238 } 1239 OS << '}'; 1240 1241 if (AllFixupsOutsideLoop) 1242 OS << ", all-fixups-outside-loop"; 1243 1244 if (WidestFixupType) 1245 OS << ", widest fixup type: " << *WidestFixupType; 1246} 1247 1248void LSRUse::dump() const { 1249 print(errs()); errs() << '\n'; 1250} 1251 1252/// isLegalUse - Test whether the use described by AM is "legal", meaning it can 1253/// be completely folded into the user instruction at isel time. This includes 1254/// address-mode folding and special icmp tricks. 1255static bool isLegalUse(const TargetLowering::AddrMode &AM, 1256 LSRUse::KindType Kind, Type *AccessTy, 1257 const TargetLowering *TLI) { 1258 switch (Kind) { 1259 case LSRUse::Address: 1260 // If we have low-level target information, ask the target if it can 1261 // completely fold this address. 1262 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy); 1263 1264 // Otherwise, just guess that reg+reg addressing is legal. 1265 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1; 1266 1267 case LSRUse::ICmpZero: 1268 // There's not even a target hook for querying whether it would be legal to 1269 // fold a GV into an ICmp. 1270 if (AM.BaseGV) 1271 return false; 1272 1273 // ICmp only has two operands; don't allow more than two non-trivial parts. 1274 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0) 1275 return false; 1276 1277 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by 1278 // putting the scaled register in the other operand of the icmp. 1279 if (AM.Scale != 0 && AM.Scale != -1) 1280 return false; 1281 1282 // If we have low-level target information, ask the target if it can fold an 1283 // integer immediate on an icmp. 1284 if (AM.BaseOffs != 0) { 1285 if (TLI) return TLI->isLegalICmpImmediate(-(uint64_t)AM.BaseOffs); 1286 return false; 1287 } 1288 1289 return true; 1290 1291 case LSRUse::Basic: 1292 // Only handle single-register values. 1293 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0; 1294 1295 case LSRUse::Special: 1296 // Only handle -1 scales, or no scale. 1297 return AM.Scale == 0 || AM.Scale == -1; 1298 } 1299 1300 llvm_unreachable("Invalid LSRUse Kind!"); 1301} 1302 1303static bool isLegalUse(TargetLowering::AddrMode AM, 1304 int64_t MinOffset, int64_t MaxOffset, 1305 LSRUse::KindType Kind, Type *AccessTy, 1306 const TargetLowering *TLI) { 1307 // Check for overflow. 1308 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) != 1309 (MinOffset > 0)) 1310 return false; 1311 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset; 1312 if (isLegalUse(AM, Kind, AccessTy, TLI)) { 1313 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset; 1314 // Check for overflow. 1315 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) != 1316 (MaxOffset > 0)) 1317 return false; 1318 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset; 1319 return isLegalUse(AM, Kind, AccessTy, TLI); 1320 } 1321 return false; 1322} 1323 1324static bool isAlwaysFoldable(int64_t BaseOffs, 1325 GlobalValue *BaseGV, 1326 bool HasBaseReg, 1327 LSRUse::KindType Kind, Type *AccessTy, 1328 const TargetLowering *TLI) { 1329 // Fast-path: zero is always foldable. 1330 if (BaseOffs == 0 && !BaseGV) return true; 1331 1332 // Conservatively, create an address with an immediate and a 1333 // base and a scale. 1334 TargetLowering::AddrMode AM; 1335 AM.BaseOffs = BaseOffs; 1336 AM.BaseGV = BaseGV; 1337 AM.HasBaseReg = HasBaseReg; 1338 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1; 1339 1340 // Canonicalize a scale of 1 to a base register if the formula doesn't 1341 // already have a base register. 1342 if (!AM.HasBaseReg && AM.Scale == 1) { 1343 AM.Scale = 0; 1344 AM.HasBaseReg = true; 1345 } 1346 1347 return isLegalUse(AM, Kind, AccessTy, TLI); 1348} 1349 1350static bool isAlwaysFoldable(const SCEV *S, 1351 int64_t MinOffset, int64_t MaxOffset, 1352 bool HasBaseReg, 1353 LSRUse::KindType Kind, Type *AccessTy, 1354 const TargetLowering *TLI, 1355 ScalarEvolution &SE) { 1356 // Fast-path: zero is always foldable. 1357 if (S->isZero()) return true; 1358 1359 // Conservatively, create an address with an immediate and a 1360 // base and a scale. 1361 int64_t BaseOffs = ExtractImmediate(S, SE); 1362 GlobalValue *BaseGV = ExtractSymbol(S, SE); 1363 1364 // If there's anything else involved, it's not foldable. 1365 if (!S->isZero()) return false; 1366 1367 // Fast-path: zero is always foldable. 1368 if (BaseOffs == 0 && !BaseGV) return true; 1369 1370 // Conservatively, create an address with an immediate and a 1371 // base and a scale. 1372 TargetLowering::AddrMode AM; 1373 AM.BaseOffs = BaseOffs; 1374 AM.BaseGV = BaseGV; 1375 AM.HasBaseReg = HasBaseReg; 1376 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1; 1377 1378 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI); 1379} 1380 1381namespace { 1382 1383/// UseMapDenseMapInfo - A DenseMapInfo implementation for holding 1384/// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind. 1385struct UseMapDenseMapInfo { 1386 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() { 1387 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic); 1388 } 1389 1390 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() { 1391 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic); 1392 } 1393 1394 static unsigned 1395 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) { 1396 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first); 1397 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second)); 1398 return Result; 1399 } 1400 1401 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS, 1402 const std::pair<const SCEV *, LSRUse::KindType> &RHS) { 1403 return LHS == RHS; 1404 } 1405}; 1406 1407/// IVInc - An individual increment in a Chain of IV increments. 1408/// Relate an IV user to an expression that computes the IV it uses from the IV 1409/// used by the previous link in the Chain. 1410/// 1411/// For the head of a chain, IncExpr holds the absolute SCEV expression for the 1412/// original IVOperand. The head of the chain's IVOperand is only valid during 1413/// chain collection, before LSR replaces IV users. During chain generation, 1414/// IncExpr can be used to find the new IVOperand that computes the same 1415/// expression. 1416struct IVInc { 1417 Instruction *UserInst; 1418 Value* IVOperand; 1419 const SCEV *IncExpr; 1420 1421 IVInc(Instruction *U, Value *O, const SCEV *E): 1422 UserInst(U), IVOperand(O), IncExpr(E) {} 1423}; 1424 1425// IVChain - The list of IV increments in program order. 1426// We typically add the head of a chain without finding subsequent links. 1427typedef SmallVector<IVInc,1> IVChain; 1428 1429/// ChainUsers - Helper for CollectChains to track multiple IV increment uses. 1430/// Distinguish between FarUsers that definitely cross IV increments and 1431/// NearUsers that may be used between IV increments. 1432struct ChainUsers { 1433 SmallPtrSet<Instruction*, 4> FarUsers; 1434 SmallPtrSet<Instruction*, 4> NearUsers; 1435}; 1436 1437/// LSRInstance - This class holds state for the main loop strength reduction 1438/// logic. 1439class LSRInstance { 1440 IVUsers &IU; 1441 ScalarEvolution &SE; 1442 DominatorTree &DT; 1443 LoopInfo &LI; 1444 const TargetLowering *const TLI; 1445 Loop *const L; 1446 bool Changed; 1447 1448 /// IVIncInsertPos - This is the insert position that the current loop's 1449 /// induction variable increment should be placed. In simple loops, this is 1450 /// the latch block's terminator. But in more complicated cases, this is a 1451 /// position which will dominate all the in-loop post-increment users. 1452 Instruction *IVIncInsertPos; 1453 1454 /// Factors - Interesting factors between use strides. 1455 SmallSetVector<int64_t, 8> Factors; 1456 1457 /// Types - Interesting use types, to facilitate truncation reuse. 1458 SmallSetVector<Type *, 4> Types; 1459 1460 /// Fixups - The list of operands which are to be replaced. 1461 SmallVector<LSRFixup, 16> Fixups; 1462 1463 /// Uses - The list of interesting uses. 1464 SmallVector<LSRUse, 16> Uses; 1465 1466 /// RegUses - Track which uses use which register candidates. 1467 RegUseTracker RegUses; 1468 1469 // Limit the number of chains to avoid quadratic behavior. We don't expect to 1470 // have more than a few IV increment chains in a loop. Missing a Chain falls 1471 // back to normal LSR behavior for those uses. 1472 static const unsigned MaxChains = 8; 1473 1474 /// IVChainVec - IV users can form a chain of IV increments. 1475 SmallVector<IVChain, MaxChains> IVChainVec; 1476 1477 /// IVIncSet - IV users that belong to profitable IVChains. 1478 SmallPtrSet<Use*, MaxChains> IVIncSet; 1479 1480 void OptimizeShadowIV(); 1481 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse); 1482 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse); 1483 void OptimizeLoopTermCond(); 1484 1485 void ChainInstruction(Instruction *UserInst, Instruction *IVOper, 1486 SmallVectorImpl<ChainUsers> &ChainUsersVec); 1487 void FinalizeChain(IVChain &Chain); 1488 void CollectChains(); 1489 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter, 1490 SmallVectorImpl<WeakVH> &DeadInsts); 1491 1492 void CollectInterestingTypesAndFactors(); 1493 void CollectFixupsAndInitialFormulae(); 1494 1495 LSRFixup &getNewFixup() { 1496 Fixups.push_back(LSRFixup()); 1497 return Fixups.back(); 1498 } 1499 1500 // Support for sharing of LSRUses between LSRFixups. 1501 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>, 1502 size_t, 1503 UseMapDenseMapInfo> UseMapTy; 1504 UseMapTy UseMap; 1505 1506 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg, 1507 LSRUse::KindType Kind, Type *AccessTy); 1508 1509 std::pair<size_t, int64_t> getUse(const SCEV *&Expr, 1510 LSRUse::KindType Kind, 1511 Type *AccessTy); 1512 1513 void DeleteUse(LSRUse &LU, size_t LUIdx); 1514 1515 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU); 1516 1517 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx); 1518 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx); 1519 void CountRegisters(const Formula &F, size_t LUIdx); 1520 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F); 1521 1522 void CollectLoopInvariantFixupsAndFormulae(); 1523 1524 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base, 1525 unsigned Depth = 0); 1526 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base); 1527 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base); 1528 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base); 1529 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base); 1530 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base); 1531 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base); 1532 void GenerateCrossUseConstantOffsets(); 1533 void GenerateAllReuseFormulae(); 1534 1535 void FilterOutUndesirableDedicatedRegisters(); 1536 1537 size_t EstimateSearchSpaceComplexity() const; 1538 void NarrowSearchSpaceByDetectingSupersets(); 1539 void NarrowSearchSpaceByCollapsingUnrolledCode(); 1540 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(); 1541 void NarrowSearchSpaceByPickingWinnerRegs(); 1542 void NarrowSearchSpaceUsingHeuristics(); 1543 1544 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution, 1545 Cost &SolutionCost, 1546 SmallVectorImpl<const Formula *> &Workspace, 1547 const Cost &CurCost, 1548 const SmallPtrSet<const SCEV *, 16> &CurRegs, 1549 DenseSet<const SCEV *> &VisitedRegs) const; 1550 void Solve(SmallVectorImpl<const Formula *> &Solution) const; 1551 1552 BasicBlock::iterator 1553 HoistInsertPosition(BasicBlock::iterator IP, 1554 const SmallVectorImpl<Instruction *> &Inputs) const; 1555 BasicBlock::iterator 1556 AdjustInsertPositionForExpand(BasicBlock::iterator IP, 1557 const LSRFixup &LF, 1558 const LSRUse &LU, 1559 SCEVExpander &Rewriter) const; 1560 1561 Value *Expand(const LSRFixup &LF, 1562 const Formula &F, 1563 BasicBlock::iterator IP, 1564 SCEVExpander &Rewriter, 1565 SmallVectorImpl<WeakVH> &DeadInsts) const; 1566 void RewriteForPHI(PHINode *PN, const LSRFixup &LF, 1567 const Formula &F, 1568 SCEVExpander &Rewriter, 1569 SmallVectorImpl<WeakVH> &DeadInsts, 1570 Pass *P) const; 1571 void Rewrite(const LSRFixup &LF, 1572 const Formula &F, 1573 SCEVExpander &Rewriter, 1574 SmallVectorImpl<WeakVH> &DeadInsts, 1575 Pass *P) const; 1576 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution, 1577 Pass *P); 1578 1579public: 1580 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P); 1581 1582 bool getChanged() const { return Changed; } 1583 1584 void print_factors_and_types(raw_ostream &OS) const; 1585 void print_fixups(raw_ostream &OS) const; 1586 void print_uses(raw_ostream &OS) const; 1587 void print(raw_ostream &OS) const; 1588 void dump() const; 1589}; 1590 1591} 1592 1593/// OptimizeShadowIV - If IV is used in a int-to-float cast 1594/// inside the loop then try to eliminate the cast operation. 1595void LSRInstance::OptimizeShadowIV() { 1596 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L); 1597 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 1598 return; 1599 1600 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); 1601 UI != E; /* empty */) { 1602 IVUsers::const_iterator CandidateUI = UI; 1603 ++UI; 1604 Instruction *ShadowUse = CandidateUI->getUser(); 1605 Type *DestTy = NULL; 1606 bool IsSigned = false; 1607 1608 /* If shadow use is a int->float cast then insert a second IV 1609 to eliminate this cast. 1610 1611 for (unsigned i = 0; i < n; ++i) 1612 foo((double)i); 1613 1614 is transformed into 1615 1616 double d = 0.0; 1617 for (unsigned i = 0; i < n; ++i, ++d) 1618 foo(d); 1619 */ 1620 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) { 1621 IsSigned = false; 1622 DestTy = UCast->getDestTy(); 1623 } 1624 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) { 1625 IsSigned = true; 1626 DestTy = SCast->getDestTy(); 1627 } 1628 if (!DestTy) continue; 1629 1630 if (TLI) { 1631 // If target does not support DestTy natively then do not apply 1632 // this transformation. 1633 EVT DVT = TLI->getValueType(DestTy); 1634 if (!TLI->isTypeLegal(DVT)) continue; 1635 } 1636 1637 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0)); 1638 if (!PH) continue; 1639 if (PH->getNumIncomingValues() != 2) continue; 1640 1641 Type *SrcTy = PH->getType(); 1642 int Mantissa = DestTy->getFPMantissaWidth(); 1643 if (Mantissa == -1) continue; 1644 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa) 1645 continue; 1646 1647 unsigned Entry, Latch; 1648 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) { 1649 Entry = 0; 1650 Latch = 1; 1651 } else { 1652 Entry = 1; 1653 Latch = 0; 1654 } 1655 1656 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry)); 1657 if (!Init) continue; 1658 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ? 1659 (double)Init->getSExtValue() : 1660 (double)Init->getZExtValue()); 1661 1662 BinaryOperator *Incr = 1663 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch)); 1664 if (!Incr) continue; 1665 if (Incr->getOpcode() != Instruction::Add 1666 && Incr->getOpcode() != Instruction::Sub) 1667 continue; 1668 1669 /* Initialize new IV, double d = 0.0 in above example. */ 1670 ConstantInt *C = NULL; 1671 if (Incr->getOperand(0) == PH) 1672 C = dyn_cast<ConstantInt>(Incr->getOperand(1)); 1673 else if (Incr->getOperand(1) == PH) 1674 C = dyn_cast<ConstantInt>(Incr->getOperand(0)); 1675 else 1676 continue; 1677 1678 if (!C) continue; 1679 1680 // Ignore negative constants, as the code below doesn't handle them 1681 // correctly. TODO: Remove this restriction. 1682 if (!C->getValue().isStrictlyPositive()) continue; 1683 1684 /* Add new PHINode. */ 1685 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH); 1686 1687 /* create new increment. '++d' in above example. */ 1688 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue()); 1689 BinaryOperator *NewIncr = 1690 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ? 1691 Instruction::FAdd : Instruction::FSub, 1692 NewPH, CFP, "IV.S.next.", Incr); 1693 1694 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry)); 1695 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch)); 1696 1697 /* Remove cast operation */ 1698 ShadowUse->replaceAllUsesWith(NewPH); 1699 ShadowUse->eraseFromParent(); 1700 Changed = true; 1701 break; 1702 } 1703} 1704 1705/// FindIVUserForCond - If Cond has an operand that is an expression of an IV, 1706/// set the IV user and stride information and return true, otherwise return 1707/// false. 1708bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) { 1709 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) 1710 if (UI->getUser() == Cond) { 1711 // NOTE: we could handle setcc instructions with multiple uses here, but 1712 // InstCombine does it as well for simple uses, it's not clear that it 1713 // occurs enough in real life to handle. 1714 CondUse = UI; 1715 return true; 1716 } 1717 return false; 1718} 1719 1720/// OptimizeMax - Rewrite the loop's terminating condition if it uses 1721/// a max computation. 1722/// 1723/// This is a narrow solution to a specific, but acute, problem. For loops 1724/// like this: 1725/// 1726/// i = 0; 1727/// do { 1728/// p[i] = 0.0; 1729/// } while (++i < n); 1730/// 1731/// the trip count isn't just 'n', because 'n' might not be positive. And 1732/// unfortunately this can come up even for loops where the user didn't use 1733/// a C do-while loop. For example, seemingly well-behaved top-test loops 1734/// will commonly be lowered like this: 1735// 1736/// if (n > 0) { 1737/// i = 0; 1738/// do { 1739/// p[i] = 0.0; 1740/// } while (++i < n); 1741/// } 1742/// 1743/// and then it's possible for subsequent optimization to obscure the if 1744/// test in such a way that indvars can't find it. 1745/// 1746/// When indvars can't find the if test in loops like this, it creates a 1747/// max expression, which allows it to give the loop a canonical 1748/// induction variable: 1749/// 1750/// i = 0; 1751/// max = n < 1 ? 1 : n; 1752/// do { 1753/// p[i] = 0.0; 1754/// } while (++i != max); 1755/// 1756/// Canonical induction variables are necessary because the loop passes 1757/// are designed around them. The most obvious example of this is the 1758/// LoopInfo analysis, which doesn't remember trip count values. It 1759/// expects to be able to rediscover the trip count each time it is 1760/// needed, and it does this using a simple analysis that only succeeds if 1761/// the loop has a canonical induction variable. 1762/// 1763/// However, when it comes time to generate code, the maximum operation 1764/// can be quite costly, especially if it's inside of an outer loop. 1765/// 1766/// This function solves this problem by detecting this type of loop and 1767/// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting 1768/// the instructions for the maximum computation. 1769/// 1770ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) { 1771 // Check that the loop matches the pattern we're looking for. 1772 if (Cond->getPredicate() != CmpInst::ICMP_EQ && 1773 Cond->getPredicate() != CmpInst::ICMP_NE) 1774 return Cond; 1775 1776 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1)); 1777 if (!Sel || !Sel->hasOneUse()) return Cond; 1778 1779 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L); 1780 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 1781 return Cond; 1782 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1); 1783 1784 // Add one to the backedge-taken count to get the trip count. 1785 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount); 1786 if (IterationCount != SE.getSCEV(Sel)) return Cond; 1787 1788 // Check for a max calculation that matches the pattern. There's no check 1789 // for ICMP_ULE here because the comparison would be with zero, which 1790 // isn't interesting. 1791 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE; 1792 const SCEVNAryExpr *Max = 0; 1793 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) { 1794 Pred = ICmpInst::ICMP_SLE; 1795 Max = S; 1796 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) { 1797 Pred = ICmpInst::ICMP_SLT; 1798 Max = S; 1799 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) { 1800 Pred = ICmpInst::ICMP_ULT; 1801 Max = U; 1802 } else { 1803 // No match; bail. 1804 return Cond; 1805 } 1806 1807 // To handle a max with more than two operands, this optimization would 1808 // require additional checking and setup. 1809 if (Max->getNumOperands() != 2) 1810 return Cond; 1811 1812 const SCEV *MaxLHS = Max->getOperand(0); 1813 const SCEV *MaxRHS = Max->getOperand(1); 1814 1815 // ScalarEvolution canonicalizes constants to the left. For < and >, look 1816 // for a comparison with 1. For <= and >=, a comparison with zero. 1817 if (!MaxLHS || 1818 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One))) 1819 return Cond; 1820 1821 // Check the relevant induction variable for conformance to 1822 // the pattern. 1823 const SCEV *IV = SE.getSCEV(Cond->getOperand(0)); 1824 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV); 1825 if (!AR || !AR->isAffine() || 1826 AR->getStart() != One || 1827 AR->getStepRecurrence(SE) != One) 1828 return Cond; 1829 1830 assert(AR->getLoop() == L && 1831 "Loop condition operand is an addrec in a different loop!"); 1832 1833 // Check the right operand of the select, and remember it, as it will 1834 // be used in the new comparison instruction. 1835 Value *NewRHS = 0; 1836 if (ICmpInst::isTrueWhenEqual(Pred)) { 1837 // Look for n+1, and grab n. 1838 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1))) 1839 if (isa<ConstantInt>(BO->getOperand(1)) && 1840 cast<ConstantInt>(BO->getOperand(1))->isOne() && 1841 SE.getSCEV(BO->getOperand(0)) == MaxRHS) 1842 NewRHS = BO->getOperand(0); 1843 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2))) 1844 if (isa<ConstantInt>(BO->getOperand(1)) && 1845 cast<ConstantInt>(BO->getOperand(1))->isOne() && 1846 SE.getSCEV(BO->getOperand(0)) == MaxRHS) 1847 NewRHS = BO->getOperand(0); 1848 if (!NewRHS) 1849 return Cond; 1850 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS) 1851 NewRHS = Sel->getOperand(1); 1852 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS) 1853 NewRHS = Sel->getOperand(2); 1854 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS)) 1855 NewRHS = SU->getValue(); 1856 else 1857 // Max doesn't match expected pattern. 1858 return Cond; 1859 1860 // Determine the new comparison opcode. It may be signed or unsigned, 1861 // and the original comparison may be either equality or inequality. 1862 if (Cond->getPredicate() == CmpInst::ICMP_EQ) 1863 Pred = CmpInst::getInversePredicate(Pred); 1864 1865 // Ok, everything looks ok to change the condition into an SLT or SGE and 1866 // delete the max calculation. 1867 ICmpInst *NewCond = 1868 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp"); 1869 1870 // Delete the max calculation instructions. 1871 Cond->replaceAllUsesWith(NewCond); 1872 CondUse->setUser(NewCond); 1873 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0)); 1874 Cond->eraseFromParent(); 1875 Sel->eraseFromParent(); 1876 if (Cmp->use_empty()) 1877 Cmp->eraseFromParent(); 1878 return NewCond; 1879} 1880 1881/// OptimizeLoopTermCond - Change loop terminating condition to use the 1882/// postinc iv when possible. 1883void 1884LSRInstance::OptimizeLoopTermCond() { 1885 SmallPtrSet<Instruction *, 4> PostIncs; 1886 1887 BasicBlock *LatchBlock = L->getLoopLatch(); 1888 SmallVector<BasicBlock*, 8> ExitingBlocks; 1889 L->getExitingBlocks(ExitingBlocks); 1890 1891 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) { 1892 BasicBlock *ExitingBlock = ExitingBlocks[i]; 1893 1894 // Get the terminating condition for the loop if possible. If we 1895 // can, we want to change it to use a post-incremented version of its 1896 // induction variable, to allow coalescing the live ranges for the IV into 1897 // one register value. 1898 1899 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator()); 1900 if (!TermBr) 1901 continue; 1902 // FIXME: Overly conservative, termination condition could be an 'or' etc.. 1903 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition())) 1904 continue; 1905 1906 // Search IVUsesByStride to find Cond's IVUse if there is one. 1907 IVStrideUse *CondUse = 0; 1908 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition()); 1909 if (!FindIVUserForCond(Cond, CondUse)) 1910 continue; 1911 1912 // If the trip count is computed in terms of a max (due to ScalarEvolution 1913 // being unable to find a sufficient guard, for example), change the loop 1914 // comparison to use SLT or ULT instead of NE. 1915 // One consequence of doing this now is that it disrupts the count-down 1916 // optimization. That's not always a bad thing though, because in such 1917 // cases it may still be worthwhile to avoid a max. 1918 Cond = OptimizeMax(Cond, CondUse); 1919 1920 // If this exiting block dominates the latch block, it may also use 1921 // the post-inc value if it won't be shared with other uses. 1922 // Check for dominance. 1923 if (!DT.dominates(ExitingBlock, LatchBlock)) 1924 continue; 1925 1926 // Conservatively avoid trying to use the post-inc value in non-latch 1927 // exits if there may be pre-inc users in intervening blocks. 1928 if (LatchBlock != ExitingBlock) 1929 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) 1930 // Test if the use is reachable from the exiting block. This dominator 1931 // query is a conservative approximation of reachability. 1932 if (&*UI != CondUse && 1933 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) { 1934 // Conservatively assume there may be reuse if the quotient of their 1935 // strides could be a legal scale. 1936 const SCEV *A = IU.getStride(*CondUse, L); 1937 const SCEV *B = IU.getStride(*UI, L); 1938 if (!A || !B) continue; 1939 if (SE.getTypeSizeInBits(A->getType()) != 1940 SE.getTypeSizeInBits(B->getType())) { 1941 if (SE.getTypeSizeInBits(A->getType()) > 1942 SE.getTypeSizeInBits(B->getType())) 1943 B = SE.getSignExtendExpr(B, A->getType()); 1944 else 1945 A = SE.getSignExtendExpr(A, B->getType()); 1946 } 1947 if (const SCEVConstant *D = 1948 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) { 1949 const ConstantInt *C = D->getValue(); 1950 // Stride of one or negative one can have reuse with non-addresses. 1951 if (C->isOne() || C->isAllOnesValue()) 1952 goto decline_post_inc; 1953 // Avoid weird situations. 1954 if (C->getValue().getMinSignedBits() >= 64 || 1955 C->getValue().isMinSignedValue()) 1956 goto decline_post_inc; 1957 // Without TLI, assume that any stride might be valid, and so any 1958 // use might be shared. 1959 if (!TLI) 1960 goto decline_post_inc; 1961 // Check for possible scaled-address reuse. 1962 Type *AccessTy = getAccessType(UI->getUser()); 1963 TargetLowering::AddrMode AM; 1964 AM.Scale = C->getSExtValue(); 1965 if (TLI->isLegalAddressingMode(AM, AccessTy)) 1966 goto decline_post_inc; 1967 AM.Scale = -AM.Scale; 1968 if (TLI->isLegalAddressingMode(AM, AccessTy)) 1969 goto decline_post_inc; 1970 } 1971 } 1972 1973 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: " 1974 << *Cond << '\n'); 1975 1976 // It's possible for the setcc instruction to be anywhere in the loop, and 1977 // possible for it to have multiple users. If it is not immediately before 1978 // the exiting block branch, move it. 1979 if (&*++BasicBlock::iterator(Cond) != TermBr) { 1980 if (Cond->hasOneUse()) { 1981 Cond->moveBefore(TermBr); 1982 } else { 1983 // Clone the terminating condition and insert into the loopend. 1984 ICmpInst *OldCond = Cond; 1985 Cond = cast<ICmpInst>(Cond->clone()); 1986 Cond->setName(L->getHeader()->getName() + ".termcond"); 1987 ExitingBlock->getInstList().insert(TermBr, Cond); 1988 1989 // Clone the IVUse, as the old use still exists! 1990 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace()); 1991 TermBr->replaceUsesOfWith(OldCond, Cond); 1992 } 1993 } 1994 1995 // If we get to here, we know that we can transform the setcc instruction to 1996 // use the post-incremented version of the IV, allowing us to coalesce the 1997 // live ranges for the IV correctly. 1998 CondUse->transformToPostInc(L); 1999 Changed = true; 2000 2001 PostIncs.insert(Cond); 2002 decline_post_inc:; 2003 } 2004 2005 // Determine an insertion point for the loop induction variable increment. It 2006 // must dominate all the post-inc comparisons we just set up, and it must 2007 // dominate the loop latch edge. 2008 IVIncInsertPos = L->getLoopLatch()->getTerminator(); 2009 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(), 2010 E = PostIncs.end(); I != E; ++I) { 2011 BasicBlock *BB = 2012 DT.findNearestCommonDominator(IVIncInsertPos->getParent(), 2013 (*I)->getParent()); 2014 if (BB == (*I)->getParent()) 2015 IVIncInsertPos = *I; 2016 else if (BB != IVIncInsertPos->getParent()) 2017 IVIncInsertPos = BB->getTerminator(); 2018 } 2019} 2020 2021/// reconcileNewOffset - Determine if the given use can accommodate a fixup 2022/// at the given offset and other details. If so, update the use and 2023/// return true. 2024bool 2025LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg, 2026 LSRUse::KindType Kind, Type *AccessTy) { 2027 int64_t NewMinOffset = LU.MinOffset; 2028 int64_t NewMaxOffset = LU.MaxOffset; 2029 Type *NewAccessTy = AccessTy; 2030 2031 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to 2032 // something conservative, however this can pessimize in the case that one of 2033 // the uses will have all its uses outside the loop, for example. 2034 if (LU.Kind != Kind) 2035 return false; 2036 // Conservatively assume HasBaseReg is true for now. 2037 if (NewOffset < LU.MinOffset) { 2038 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg, 2039 Kind, AccessTy, TLI)) 2040 return false; 2041 NewMinOffset = NewOffset; 2042 } else if (NewOffset > LU.MaxOffset) { 2043 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg, 2044 Kind, AccessTy, TLI)) 2045 return false; 2046 NewMaxOffset = NewOffset; 2047 } 2048 // Check for a mismatched access type, and fall back conservatively as needed. 2049 // TODO: Be less conservative when the type is similar and can use the same 2050 // addressing modes. 2051 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy) 2052 NewAccessTy = Type::getVoidTy(AccessTy->getContext()); 2053 2054 // Update the use. 2055 LU.MinOffset = NewMinOffset; 2056 LU.MaxOffset = NewMaxOffset; 2057 LU.AccessTy = NewAccessTy; 2058 if (NewOffset != LU.Offsets.back()) 2059 LU.Offsets.push_back(NewOffset); 2060 return true; 2061} 2062 2063/// getUse - Return an LSRUse index and an offset value for a fixup which 2064/// needs the given expression, with the given kind and optional access type. 2065/// Either reuse an existing use or create a new one, as needed. 2066std::pair<size_t, int64_t> 2067LSRInstance::getUse(const SCEV *&Expr, 2068 LSRUse::KindType Kind, Type *AccessTy) { 2069 const SCEV *Copy = Expr; 2070 int64_t Offset = ExtractImmediate(Expr, SE); 2071 2072 // Basic uses can't accept any offset, for example. 2073 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) { 2074 Expr = Copy; 2075 Offset = 0; 2076 } 2077 2078 std::pair<UseMapTy::iterator, bool> P = 2079 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0)); 2080 if (!P.second) { 2081 // A use already existed with this base. 2082 size_t LUIdx = P.first->second; 2083 LSRUse &LU = Uses[LUIdx]; 2084 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy)) 2085 // Reuse this use. 2086 return std::make_pair(LUIdx, Offset); 2087 } 2088 2089 // Create a new use. 2090 size_t LUIdx = Uses.size(); 2091 P.first->second = LUIdx; 2092 Uses.push_back(LSRUse(Kind, AccessTy)); 2093 LSRUse &LU = Uses[LUIdx]; 2094 2095 // We don't need to track redundant offsets, but we don't need to go out 2096 // of our way here to avoid them. 2097 if (LU.Offsets.empty() || Offset != LU.Offsets.back()) 2098 LU.Offsets.push_back(Offset); 2099 2100 LU.MinOffset = Offset; 2101 LU.MaxOffset = Offset; 2102 return std::make_pair(LUIdx, Offset); 2103} 2104 2105/// DeleteUse - Delete the given use from the Uses list. 2106void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) { 2107 if (&LU != &Uses.back()) 2108 std::swap(LU, Uses.back()); 2109 Uses.pop_back(); 2110 2111 // Update RegUses. 2112 RegUses.SwapAndDropUse(LUIdx, Uses.size()); 2113} 2114 2115/// FindUseWithFormula - Look for a use distinct from OrigLU which is has 2116/// a formula that has the same registers as the given formula. 2117LSRUse * 2118LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF, 2119 const LSRUse &OrigLU) { 2120 // Search all uses for the formula. This could be more clever. 2121 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 2122 LSRUse &LU = Uses[LUIdx]; 2123 // Check whether this use is close enough to OrigLU, to see whether it's 2124 // worthwhile looking through its formulae. 2125 // Ignore ICmpZero uses because they may contain formulae generated by 2126 // GenerateICmpZeroScales, in which case adding fixup offsets may 2127 // be invalid. 2128 if (&LU != &OrigLU && 2129 LU.Kind != LSRUse::ICmpZero && 2130 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy && 2131 LU.WidestFixupType == OrigLU.WidestFixupType && 2132 LU.HasFormulaWithSameRegs(OrigF)) { 2133 // Scan through this use's formulae. 2134 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(), 2135 E = LU.Formulae.end(); I != E; ++I) { 2136 const Formula &F = *I; 2137 // Check to see if this formula has the same registers and symbols 2138 // as OrigF. 2139 if (F.BaseRegs == OrigF.BaseRegs && 2140 F.ScaledReg == OrigF.ScaledReg && 2141 F.AM.BaseGV == OrigF.AM.BaseGV && 2142 F.AM.Scale == OrigF.AM.Scale && 2143 F.UnfoldedOffset == OrigF.UnfoldedOffset) { 2144 if (F.AM.BaseOffs == 0) 2145 return &LU; 2146 // This is the formula where all the registers and symbols matched; 2147 // there aren't going to be any others. Since we declined it, we 2148 // can skip the rest of the formulae and procede to the next LSRUse. 2149 break; 2150 } 2151 } 2152 } 2153 } 2154 2155 // Nothing looked good. 2156 return 0; 2157} 2158 2159void LSRInstance::CollectInterestingTypesAndFactors() { 2160 SmallSetVector<const SCEV *, 4> Strides; 2161 2162 // Collect interesting types and strides. 2163 SmallVector<const SCEV *, 4> Worklist; 2164 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) { 2165 const SCEV *Expr = IU.getExpr(*UI); 2166 2167 // Collect interesting types. 2168 Types.insert(SE.getEffectiveSCEVType(Expr->getType())); 2169 2170 // Add strides for mentioned loops. 2171 Worklist.push_back(Expr); 2172 do { 2173 const SCEV *S = Worklist.pop_back_val(); 2174 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 2175 if (AR->getLoop() == L) 2176 Strides.insert(AR->getStepRecurrence(SE)); 2177 Worklist.push_back(AR->getStart()); 2178 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 2179 Worklist.append(Add->op_begin(), Add->op_end()); 2180 } 2181 } while (!Worklist.empty()); 2182 } 2183 2184 // Compute interesting factors from the set of interesting strides. 2185 for (SmallSetVector<const SCEV *, 4>::const_iterator 2186 I = Strides.begin(), E = Strides.end(); I != E; ++I) 2187 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter = 2188 llvm::next(I); NewStrideIter != E; ++NewStrideIter) { 2189 const SCEV *OldStride = *I; 2190 const SCEV *NewStride = *NewStrideIter; 2191 2192 if (SE.getTypeSizeInBits(OldStride->getType()) != 2193 SE.getTypeSizeInBits(NewStride->getType())) { 2194 if (SE.getTypeSizeInBits(OldStride->getType()) > 2195 SE.getTypeSizeInBits(NewStride->getType())) 2196 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType()); 2197 else 2198 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType()); 2199 } 2200 if (const SCEVConstant *Factor = 2201 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride, 2202 SE, true))) { 2203 if (Factor->getValue()->getValue().getMinSignedBits() <= 64) 2204 Factors.insert(Factor->getValue()->getValue().getSExtValue()); 2205 } else if (const SCEVConstant *Factor = 2206 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride, 2207 NewStride, 2208 SE, true))) { 2209 if (Factor->getValue()->getValue().getMinSignedBits() <= 64) 2210 Factors.insert(Factor->getValue()->getValue().getSExtValue()); 2211 } 2212 } 2213 2214 // If all uses use the same type, don't bother looking for truncation-based 2215 // reuse. 2216 if (Types.size() == 1) 2217 Types.clear(); 2218 2219 DEBUG(print_factors_and_types(dbgs())); 2220} 2221 2222/// findIVOperand - Helper for CollectChains that finds an IV operand (computed 2223/// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped 2224/// Instructions to IVStrideUses, we could partially skip this. 2225static User::op_iterator 2226findIVOperand(User::op_iterator OI, User::op_iterator OE, 2227 Loop *L, ScalarEvolution &SE) { 2228 for(; OI != OE; ++OI) { 2229 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) { 2230 if (!SE.isSCEVable(Oper->getType())) 2231 continue; 2232 2233 if (const SCEVAddRecExpr *AR = 2234 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) { 2235 if (AR->getLoop() == L) 2236 break; 2237 } 2238 } 2239 } 2240 return OI; 2241} 2242 2243/// getWideOperand - IVChain logic must consistenctly peek base TruncInst 2244/// operands, so wrap it in a convenient helper. 2245static Value *getWideOperand(Value *Oper) { 2246 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper)) 2247 return Trunc->getOperand(0); 2248 return Oper; 2249} 2250 2251/// isCompatibleIVType - Return true if we allow an IV chain to include both 2252/// types. 2253static bool isCompatibleIVType(Value *LVal, Value *RVal) { 2254 Type *LType = LVal->getType(); 2255 Type *RType = RVal->getType(); 2256 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy()); 2257} 2258 2259/// getExprBase - Return an approximation of this SCEV expression's "base", or 2260/// NULL for any constant. Returning the expression itself is 2261/// conservative. Returning a deeper subexpression is more precise and valid as 2262/// long as it isn't less complex than another subexpression. For expressions 2263/// involving multiple unscaled values, we need to return the pointer-type 2264/// SCEVUnknown. This avoids forming chains across objects, such as: 2265/// PrevOper==a[i], IVOper==b[i], IVInc==b-a. 2266/// 2267/// Since SCEVUnknown is the rightmost type, and pointers are the rightmost 2268/// SCEVUnknown, we simply return the rightmost SCEV operand. 2269static const SCEV *getExprBase(const SCEV *S) { 2270 switch (S->getSCEVType()) { 2271 default: // uncluding scUnknown. 2272 return S; 2273 case scConstant: 2274 return 0; 2275 case scTruncate: 2276 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand()); 2277 case scZeroExtend: 2278 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand()); 2279 case scSignExtend: 2280 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand()); 2281 case scAddExpr: { 2282 // Skip over scaled operands (scMulExpr) to follow add operands as long as 2283 // there's nothing more complex. 2284 // FIXME: not sure if we want to recognize negation. 2285 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S); 2286 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()), 2287 E(Add->op_begin()); I != E; ++I) { 2288 const SCEV *SubExpr = *I; 2289 if (SubExpr->getSCEVType() == scAddExpr) 2290 return getExprBase(SubExpr); 2291 2292 if (SubExpr->getSCEVType() != scMulExpr) 2293 return SubExpr; 2294 } 2295 return S; // all operands are scaled, be conservative. 2296 } 2297 case scAddRecExpr: 2298 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart()); 2299 } 2300} 2301 2302/// Return true if the chain increment is profitable to expand into a loop 2303/// invariant value, which may require its own register. A profitable chain 2304/// increment will be an offset relative to the same base. We allow such offsets 2305/// to potentially be used as chain increment as long as it's not obviously 2306/// expensive to expand using real instructions. 2307static const SCEV * 2308getProfitableChainIncrement(Value *NextIV, Value *PrevIV, 2309 const IVChain &Chain, Loop *L, 2310 ScalarEvolution &SE, const TargetLowering *TLI) { 2311 // Prune the solution space aggressively by checking that both IV operands 2312 // are expressions that operate on the same unscaled SCEVUnknown. This 2313 // "base" will be canceled by the subsequent getMinusSCEV call. Checking first 2314 // avoids creating extra SCEV expressions. 2315 const SCEV *OperExpr = SE.getSCEV(NextIV); 2316 const SCEV *PrevExpr = SE.getSCEV(PrevIV); 2317 if (getExprBase(OperExpr) != getExprBase(PrevExpr) && !StressIVChain) 2318 return 0; 2319 2320 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr); 2321 if (!SE.isLoopInvariant(IncExpr, L)) 2322 return 0; 2323 2324 // We are not able to expand an increment unless it is loop invariant, 2325 // however, the following checks are purely for profitability. 2326 if (StressIVChain) 2327 return IncExpr; 2328 2329 // Do not replace a constant offset from IV head with a nonconstant IV 2330 // increment. 2331 if (!isa<SCEVConstant>(IncExpr)) { 2332 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Chain[0].IVOperand)); 2333 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr))) 2334 return 0; 2335 } 2336 2337 SmallPtrSet<const SCEV*, 8> Processed; 2338 if (isHighCostExpansion(IncExpr, Processed, SE)) 2339 return 0; 2340 2341 return IncExpr; 2342} 2343 2344/// Return true if the number of registers needed for the chain is estimated to 2345/// be less than the number required for the individual IV users. First prohibit 2346/// any IV users that keep the IV live across increments (the Users set should 2347/// be empty). Next count the number and type of increments in the chain. 2348/// 2349/// Chaining IVs can lead to considerable code bloat if ISEL doesn't 2350/// effectively use postinc addressing modes. Only consider it profitable it the 2351/// increments can be computed in fewer registers when chained. 2352/// 2353/// TODO: Consider IVInc free if it's already used in another chains. 2354static bool 2355isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users, 2356 ScalarEvolution &SE, const TargetLowering *TLI) { 2357 if (StressIVChain) 2358 return true; 2359 2360 if (Chain.size() <= 2) 2361 return false; 2362 2363 if (!Users.empty()) { 2364 DEBUG(dbgs() << "Chain: " << *Chain[0].UserInst << " users:\n"; 2365 for (SmallPtrSet<Instruction*, 4>::const_iterator I = Users.begin(), 2366 E = Users.end(); I != E; ++I) { 2367 dbgs() << " " << **I << "\n"; 2368 }); 2369 return false; 2370 } 2371 assert(!Chain.empty() && "empty IV chains are not allowed"); 2372 2373 // The chain itself may require a register, so intialize cost to 1. 2374 int cost = 1; 2375 2376 // A complete chain likely eliminates the need for keeping the original IV in 2377 // a register. LSR does not currently know how to form a complete chain unless 2378 // the header phi already exists. 2379 if (isa<PHINode>(Chain.back().UserInst) 2380 && SE.getSCEV(Chain.back().UserInst) == Chain[0].IncExpr) { 2381 --cost; 2382 } 2383 const SCEV *LastIncExpr = 0; 2384 unsigned NumConstIncrements = 0; 2385 unsigned NumVarIncrements = 0; 2386 unsigned NumReusedIncrements = 0; 2387 for (IVChain::const_iterator I = llvm::next(Chain.begin()), E = Chain.end(); 2388 I != E; ++I) { 2389 2390 if (I->IncExpr->isZero()) 2391 continue; 2392 2393 // Incrementing by zero or some constant is neutral. We assume constants can 2394 // be folded into an addressing mode or an add's immediate operand. 2395 if (isa<SCEVConstant>(I->IncExpr)) { 2396 ++NumConstIncrements; 2397 continue; 2398 } 2399 2400 if (I->IncExpr == LastIncExpr) 2401 ++NumReusedIncrements; 2402 else 2403 ++NumVarIncrements; 2404 2405 LastIncExpr = I->IncExpr; 2406 } 2407 // An IV chain with a single increment is handled by LSR's postinc 2408 // uses. However, a chain with multiple increments requires keeping the IV's 2409 // value live longer than it needs to be if chained. 2410 if (NumConstIncrements > 1) 2411 --cost; 2412 2413 // Materializing increment expressions in the preheader that didn't exist in 2414 // the original code may cost a register. For example, sign-extended array 2415 // indices can produce ridiculous increments like this: 2416 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64))) 2417 cost += NumVarIncrements; 2418 2419 // Reusing variable increments likely saves a register to hold the multiple of 2420 // the stride. 2421 cost -= NumReusedIncrements; 2422 2423 DEBUG(dbgs() << "Chain: " << *Chain[0].UserInst << " Cost: " << cost << "\n"); 2424 2425 return cost < 0; 2426} 2427 2428/// ChainInstruction - Add this IV user to an existing chain or make it the head 2429/// of a new chain. 2430void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper, 2431 SmallVectorImpl<ChainUsers> &ChainUsersVec) { 2432 // When IVs are used as types of varying widths, they are generally converted 2433 // to a wider type with some uses remaining narrow under a (free) trunc. 2434 Value *NextIV = getWideOperand(IVOper); 2435 2436 // Visit all existing chains. Check if its IVOper can be computed as a 2437 // profitable loop invariant increment from the last link in the Chain. 2438 unsigned ChainIdx = 0, NChains = IVChainVec.size(); 2439 const SCEV *LastIncExpr = 0; 2440 for (; ChainIdx < NChains; ++ChainIdx) { 2441 Value *PrevIV = getWideOperand(IVChainVec[ChainIdx].back().IVOperand); 2442 if (!isCompatibleIVType(PrevIV, NextIV)) 2443 continue; 2444 2445 // A phi node terminates a chain. 2446 if (isa<PHINode>(UserInst) 2447 && isa<PHINode>(IVChainVec[ChainIdx].back().UserInst)) 2448 continue; 2449 2450 if (const SCEV *IncExpr = 2451 getProfitableChainIncrement(NextIV, PrevIV, IVChainVec[ChainIdx], 2452 L, SE, TLI)) { 2453 LastIncExpr = IncExpr; 2454 break; 2455 } 2456 } 2457 // If we haven't found a chain, create a new one, unless we hit the max. Don't 2458 // bother for phi nodes, because they must be last in the chain. 2459 if (ChainIdx == NChains) { 2460 if (isa<PHINode>(UserInst)) 2461 return; 2462 if (NChains >= MaxChains && !StressIVChain) { 2463 DEBUG(dbgs() << "IV Chain Limit\n"); 2464 return; 2465 } 2466 LastIncExpr = SE.getSCEV(NextIV); 2467 // IVUsers may have skipped over sign/zero extensions. We don't currently 2468 // attempt to form chains involving extensions unless they can be hoisted 2469 // into this loop's AddRec. 2470 if (!isa<SCEVAddRecExpr>(LastIncExpr)) 2471 return; 2472 ++NChains; 2473 IVChainVec.resize(NChains); 2474 ChainUsersVec.resize(NChains); 2475 DEBUG(dbgs() << "IV Head: (" << *UserInst << ") IV=" << *LastIncExpr 2476 << "\n"); 2477 } 2478 else 2479 DEBUG(dbgs() << "IV Inc: (" << *UserInst << ") IV+" << *LastIncExpr 2480 << "\n"); 2481 2482 // Add this IV user to the end of the chain. 2483 IVChainVec[ChainIdx].push_back(IVInc(UserInst, IVOper, LastIncExpr)); 2484 2485 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers; 2486 // This chain's NearUsers become FarUsers. 2487 if (!LastIncExpr->isZero()) { 2488 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(), 2489 NearUsers.end()); 2490 NearUsers.clear(); 2491 } 2492 2493 // All other uses of IVOperand become near uses of the chain. 2494 // We currently ignore intermediate values within SCEV expressions, assuming 2495 // they will eventually be used be the current chain, or can be computed 2496 // from one of the chain increments. To be more precise we could 2497 // transitively follow its user and only add leaf IV users to the set. 2498 for (Value::use_iterator UseIter = IVOper->use_begin(), 2499 UseEnd = IVOper->use_end(); UseIter != UseEnd; ++UseIter) { 2500 Instruction *OtherUse = dyn_cast<Instruction>(*UseIter); 2501 if (!OtherUse || OtherUse == UserInst) 2502 continue; 2503 if (SE.isSCEVable(OtherUse->getType()) 2504 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse)) 2505 && IU.isIVUserOrOperand(OtherUse)) { 2506 continue; 2507 } 2508 NearUsers.insert(OtherUse); 2509 } 2510 2511 // Since this user is part of the chain, it's no longer considered a use 2512 // of the chain. 2513 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst); 2514} 2515 2516/// CollectChains - Populate the vector of Chains. 2517/// 2518/// This decreases ILP at the architecture level. Targets with ample registers, 2519/// multiple memory ports, and no register renaming probably don't want 2520/// this. However, such targets should probably disable LSR altogether. 2521/// 2522/// The job of LSR is to make a reasonable choice of induction variables across 2523/// the loop. Subsequent passes can easily "unchain" computation exposing more 2524/// ILP *within the loop* if the target wants it. 2525/// 2526/// Finding the best IV chain is potentially a scheduling problem. Since LSR 2527/// will not reorder memory operations, it will recognize this as a chain, but 2528/// will generate redundant IV increments. Ideally this would be corrected later 2529/// by a smart scheduler: 2530/// = A[i] 2531/// = A[i+x] 2532/// A[i] = 2533/// A[i+x] = 2534/// 2535/// TODO: Walk the entire domtree within this loop, not just the path to the 2536/// loop latch. This will discover chains on side paths, but requires 2537/// maintaining multiple copies of the Chains state. 2538void LSRInstance::CollectChains() { 2539 SmallVector<ChainUsers, 8> ChainUsersVec; 2540 2541 SmallVector<BasicBlock *,8> LatchPath; 2542 BasicBlock *LoopHeader = L->getHeader(); 2543 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch()); 2544 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) { 2545 LatchPath.push_back(Rung->getBlock()); 2546 } 2547 LatchPath.push_back(LoopHeader); 2548 2549 // Walk the instruction stream from the loop header to the loop latch. 2550 for (SmallVectorImpl<BasicBlock *>::reverse_iterator 2551 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend(); 2552 BBIter != BBEnd; ++BBIter) { 2553 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end(); 2554 I != E; ++I) { 2555 // Skip instructions that weren't seen by IVUsers analysis. 2556 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I)) 2557 continue; 2558 2559 // Ignore users that are part of a SCEV expression. This way we only 2560 // consider leaf IV Users. This effectively rediscovers a portion of 2561 // IVUsers analysis but in program order this time. 2562 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I))) 2563 continue; 2564 2565 // Remove this instruction from any NearUsers set it may be in. 2566 for (unsigned ChainIdx = 0, NChains = IVChainVec.size(); 2567 ChainIdx < NChains; ++ChainIdx) { 2568 ChainUsersVec[ChainIdx].NearUsers.erase(I); 2569 } 2570 // Search for operands that can be chained. 2571 SmallPtrSet<Instruction*, 4> UniqueOperands; 2572 User::op_iterator IVOpEnd = I->op_end(); 2573 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE); 2574 while (IVOpIter != IVOpEnd) { 2575 Instruction *IVOpInst = cast<Instruction>(*IVOpIter); 2576 if (UniqueOperands.insert(IVOpInst)) 2577 ChainInstruction(I, IVOpInst, ChainUsersVec); 2578 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE); 2579 } 2580 } // Continue walking down the instructions. 2581 } // Continue walking down the domtree. 2582 // Visit phi backedges to determine if the chain can generate the IV postinc. 2583 for (BasicBlock::iterator I = L->getHeader()->begin(); 2584 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 2585 if (!SE.isSCEVable(PN->getType())) 2586 continue; 2587 2588 Instruction *IncV = 2589 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch())); 2590 if (IncV) 2591 ChainInstruction(PN, IncV, ChainUsersVec); 2592 } 2593 // Remove any unprofitable chains. 2594 unsigned ChainIdx = 0; 2595 for (unsigned UsersIdx = 0, NChains = IVChainVec.size(); 2596 UsersIdx < NChains; ++UsersIdx) { 2597 if (!isProfitableChain(IVChainVec[UsersIdx], 2598 ChainUsersVec[UsersIdx].FarUsers, SE, TLI)) 2599 continue; 2600 // Preserve the chain at UsesIdx. 2601 if (ChainIdx != UsersIdx) 2602 IVChainVec[ChainIdx] = IVChainVec[UsersIdx]; 2603 FinalizeChain(IVChainVec[ChainIdx]); 2604 ++ChainIdx; 2605 } 2606 IVChainVec.resize(ChainIdx); 2607} 2608 2609void LSRInstance::FinalizeChain(IVChain &Chain) { 2610 assert(!Chain.empty() && "empty IV chains are not allowed"); 2611 DEBUG(dbgs() << "Final Chain: " << *Chain[0].UserInst << "\n"); 2612 2613 for (IVChain::const_iterator I = llvm::next(Chain.begin()), E = Chain.end(); 2614 I != E; ++I) { 2615 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n"); 2616 User::op_iterator UseI = 2617 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand); 2618 assert(UseI != I->UserInst->op_end() && "cannot find IV operand"); 2619 IVIncSet.insert(UseI); 2620 } 2621} 2622 2623/// Return true if the IVInc can be folded into an addressing mode. 2624static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst, 2625 Value *Operand, const TargetLowering *TLI) { 2626 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr); 2627 if (!IncConst || !isAddressUse(UserInst, Operand)) 2628 return false; 2629 2630 if (IncConst->getValue()->getValue().getMinSignedBits() > 64) 2631 return false; 2632 2633 int64_t IncOffset = IncConst->getValue()->getSExtValue(); 2634 if (!isAlwaysFoldable(IncOffset, /*BaseGV=*/0, /*HaseBaseReg=*/false, 2635 LSRUse::Address, getAccessType(UserInst), TLI)) 2636 return false; 2637 2638 return true; 2639} 2640 2641/// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to 2642/// materialize the IV user's operand from the previous IV user's operand. 2643void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter, 2644 SmallVectorImpl<WeakVH> &DeadInsts) { 2645 // Find the new IVOperand for the head of the chain. It may have been replaced 2646 // by LSR. 2647 const IVInc &Head = Chain[0]; 2648 User::op_iterator IVOpEnd = Head.UserInst->op_end(); 2649 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(), 2650 IVOpEnd, L, SE); 2651 Value *IVSrc = 0; 2652 while (IVOpIter != IVOpEnd) { 2653 IVSrc = getWideOperand(*IVOpIter); 2654 2655 // If this operand computes the expression that the chain needs, we may use 2656 // it. (Check this after setting IVSrc which is used below.) 2657 // 2658 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too 2659 // narrow for the chain, so we can no longer use it. We do allow using a 2660 // wider phi, assuming the LSR checked for free truncation. In that case we 2661 // should already have a truncate on this operand such that 2662 // getSCEV(IVSrc) == IncExpr. 2663 if (SE.getSCEV(*IVOpIter) == Head.IncExpr 2664 || SE.getSCEV(IVSrc) == Head.IncExpr) { 2665 break; 2666 } 2667 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE); 2668 } 2669 if (IVOpIter == IVOpEnd) { 2670 // Gracefully give up on this chain. 2671 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n"); 2672 return; 2673 } 2674 2675 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n"); 2676 Type *IVTy = IVSrc->getType(); 2677 Type *IntTy = SE.getEffectiveSCEVType(IVTy); 2678 const SCEV *LeftOverExpr = 0; 2679 for (IVChain::const_iterator IncI = llvm::next(Chain.begin()), 2680 IncE = Chain.end(); IncI != IncE; ++IncI) { 2681 2682 Instruction *InsertPt = IncI->UserInst; 2683 if (isa<PHINode>(InsertPt)) 2684 InsertPt = L->getLoopLatch()->getTerminator(); 2685 2686 // IVOper will replace the current IV User's operand. IVSrc is the IV 2687 // value currently held in a register. 2688 Value *IVOper = IVSrc; 2689 if (!IncI->IncExpr->isZero()) { 2690 // IncExpr was the result of subtraction of two narrow values, so must 2691 // be signed. 2692 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy); 2693 LeftOverExpr = LeftOverExpr ? 2694 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr; 2695 } 2696 if (LeftOverExpr && !LeftOverExpr->isZero()) { 2697 // Expand the IV increment. 2698 Rewriter.clearPostInc(); 2699 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt); 2700 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc), 2701 SE.getUnknown(IncV)); 2702 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt); 2703 2704 // If an IV increment can't be folded, use it as the next IV value. 2705 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand, 2706 TLI)) { 2707 assert(IVTy == IVOper->getType() && "inconsistent IV increment type"); 2708 IVSrc = IVOper; 2709 LeftOverExpr = 0; 2710 } 2711 } 2712 Type *OperTy = IncI->IVOperand->getType(); 2713 if (IVTy != OperTy) { 2714 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) && 2715 "cannot extend a chained IV"); 2716 IRBuilder<> Builder(InsertPt); 2717 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain"); 2718 } 2719 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper); 2720 DeadInsts.push_back(IncI->IVOperand); 2721 } 2722 // If LSR created a new, wider phi, we may also replace its postinc. We only 2723 // do this if we also found a wide value for the head of the chain. 2724 if (isa<PHINode>(Chain.back().UserInst)) { 2725 for (BasicBlock::iterator I = L->getHeader()->begin(); 2726 PHINode *Phi = dyn_cast<PHINode>(I); ++I) { 2727 if (!isCompatibleIVType(Phi, IVSrc)) 2728 continue; 2729 Instruction *PostIncV = dyn_cast<Instruction>( 2730 Phi->getIncomingValueForBlock(L->getLoopLatch())); 2731 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc))) 2732 continue; 2733 Value *IVOper = IVSrc; 2734 Type *PostIncTy = PostIncV->getType(); 2735 if (IVTy != PostIncTy) { 2736 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types"); 2737 IRBuilder<> Builder(L->getLoopLatch()->getTerminator()); 2738 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc()); 2739 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain"); 2740 } 2741 Phi->replaceUsesOfWith(PostIncV, IVOper); 2742 DeadInsts.push_back(PostIncV); 2743 } 2744 } 2745} 2746 2747void LSRInstance::CollectFixupsAndInitialFormulae() { 2748 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) { 2749 Instruction *UserInst = UI->getUser(); 2750 // Skip IV users that are part of profitable IV Chains. 2751 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(), 2752 UI->getOperandValToReplace()); 2753 assert(UseI != UserInst->op_end() && "cannot find IV operand"); 2754 if (IVIncSet.count(UseI)) 2755 continue; 2756 2757 // Record the uses. 2758 LSRFixup &LF = getNewFixup(); 2759 LF.UserInst = UserInst; 2760 LF.OperandValToReplace = UI->getOperandValToReplace(); 2761 LF.PostIncLoops = UI->getPostIncLoops(); 2762 2763 LSRUse::KindType Kind = LSRUse::Basic; 2764 Type *AccessTy = 0; 2765 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) { 2766 Kind = LSRUse::Address; 2767 AccessTy = getAccessType(LF.UserInst); 2768 } 2769 2770 const SCEV *S = IU.getExpr(*UI); 2771 2772 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as 2773 // (N - i == 0), and this allows (N - i) to be the expression that we work 2774 // with rather than just N or i, so we can consider the register 2775 // requirements for both N and i at the same time. Limiting this code to 2776 // equality icmps is not a problem because all interesting loops use 2777 // equality icmps, thanks to IndVarSimplify. 2778 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst)) 2779 if (CI->isEquality()) { 2780 // Swap the operands if needed to put the OperandValToReplace on the 2781 // left, for consistency. 2782 Value *NV = CI->getOperand(1); 2783 if (NV == LF.OperandValToReplace) { 2784 CI->setOperand(1, CI->getOperand(0)); 2785 CI->setOperand(0, NV); 2786 NV = CI->getOperand(1); 2787 Changed = true; 2788 } 2789 2790 // x == y --> x - y == 0 2791 const SCEV *N = SE.getSCEV(NV); 2792 if (SE.isLoopInvariant(N, L)) { 2793 // S is normalized, so normalize N before folding it into S 2794 // to keep the result normalized. 2795 N = TransformForPostIncUse(Normalize, N, CI, 0, 2796 LF.PostIncLoops, SE, DT); 2797 Kind = LSRUse::ICmpZero; 2798 S = SE.getMinusSCEV(N, S); 2799 } 2800 2801 // -1 and the negations of all interesting strides (except the negation 2802 // of -1) are now also interesting. 2803 for (size_t i = 0, e = Factors.size(); i != e; ++i) 2804 if (Factors[i] != -1) 2805 Factors.insert(-(uint64_t)Factors[i]); 2806 Factors.insert(-1); 2807 } 2808 2809 // Set up the initial formula for this use. 2810 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy); 2811 LF.LUIdx = P.first; 2812 LF.Offset = P.second; 2813 LSRUse &LU = Uses[LF.LUIdx]; 2814 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L); 2815 if (!LU.WidestFixupType || 2816 SE.getTypeSizeInBits(LU.WidestFixupType) < 2817 SE.getTypeSizeInBits(LF.OperandValToReplace->getType())) 2818 LU.WidestFixupType = LF.OperandValToReplace->getType(); 2819 2820 // If this is the first use of this LSRUse, give it a formula. 2821 if (LU.Formulae.empty()) { 2822 InsertInitialFormula(S, LU, LF.LUIdx); 2823 CountRegisters(LU.Formulae.back(), LF.LUIdx); 2824 } 2825 } 2826 2827 DEBUG(print_fixups(dbgs())); 2828} 2829 2830/// InsertInitialFormula - Insert a formula for the given expression into 2831/// the given use, separating out loop-variant portions from loop-invariant 2832/// and loop-computable portions. 2833void 2834LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) { 2835 Formula F; 2836 F.InitialMatch(S, L, SE); 2837 bool Inserted = InsertFormula(LU, LUIdx, F); 2838 assert(Inserted && "Initial formula already exists!"); (void)Inserted; 2839} 2840 2841/// InsertSupplementalFormula - Insert a simple single-register formula for 2842/// the given expression into the given use. 2843void 2844LSRInstance::InsertSupplementalFormula(const SCEV *S, 2845 LSRUse &LU, size_t LUIdx) { 2846 Formula F; 2847 F.BaseRegs.push_back(S); 2848 F.AM.HasBaseReg = true; 2849 bool Inserted = InsertFormula(LU, LUIdx, F); 2850 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted; 2851} 2852 2853/// CountRegisters - Note which registers are used by the given formula, 2854/// updating RegUses. 2855void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) { 2856 if (F.ScaledReg) 2857 RegUses.CountRegister(F.ScaledReg, LUIdx); 2858 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(), 2859 E = F.BaseRegs.end(); I != E; ++I) 2860 RegUses.CountRegister(*I, LUIdx); 2861} 2862 2863/// InsertFormula - If the given formula has not yet been inserted, add it to 2864/// the list, and return true. Return false otherwise. 2865bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) { 2866 if (!LU.InsertFormula(F)) 2867 return false; 2868 2869 CountRegisters(F, LUIdx); 2870 return true; 2871} 2872 2873/// CollectLoopInvariantFixupsAndFormulae - Check for other uses of 2874/// loop-invariant values which we're tracking. These other uses will pin these 2875/// values in registers, making them less profitable for elimination. 2876/// TODO: This currently misses non-constant addrec step registers. 2877/// TODO: Should this give more weight to users inside the loop? 2878void 2879LSRInstance::CollectLoopInvariantFixupsAndFormulae() { 2880 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end()); 2881 SmallPtrSet<const SCEV *, 8> Inserted; 2882 2883 while (!Worklist.empty()) { 2884 const SCEV *S = Worklist.pop_back_val(); 2885 2886 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) 2887 Worklist.append(N->op_begin(), N->op_end()); 2888 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) 2889 Worklist.push_back(C->getOperand()); 2890 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 2891 Worklist.push_back(D->getLHS()); 2892 Worklist.push_back(D->getRHS()); 2893 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 2894 if (!Inserted.insert(U)) continue; 2895 const Value *V = U->getValue(); 2896 if (const Instruction *Inst = dyn_cast<Instruction>(V)) { 2897 // Look for instructions defined outside the loop. 2898 if (L->contains(Inst)) continue; 2899 } else if (isa<UndefValue>(V)) 2900 // Undef doesn't have a live range, so it doesn't matter. 2901 continue; 2902 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end(); 2903 UI != UE; ++UI) { 2904 const Instruction *UserInst = dyn_cast<Instruction>(*UI); 2905 // Ignore non-instructions. 2906 if (!UserInst) 2907 continue; 2908 // Ignore instructions in other functions (as can happen with 2909 // Constants). 2910 if (UserInst->getParent()->getParent() != L->getHeader()->getParent()) 2911 continue; 2912 // Ignore instructions not dominated by the loop. 2913 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ? 2914 UserInst->getParent() : 2915 cast<PHINode>(UserInst)->getIncomingBlock( 2916 PHINode::getIncomingValueNumForOperand(UI.getOperandNo())); 2917 if (!DT.dominates(L->getHeader(), UseBB)) 2918 continue; 2919 // Ignore uses which are part of other SCEV expressions, to avoid 2920 // analyzing them multiple times. 2921 if (SE.isSCEVable(UserInst->getType())) { 2922 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst)); 2923 // If the user is a no-op, look through to its uses. 2924 if (!isa<SCEVUnknown>(UserS)) 2925 continue; 2926 if (UserS == U) { 2927 Worklist.push_back( 2928 SE.getUnknown(const_cast<Instruction *>(UserInst))); 2929 continue; 2930 } 2931 } 2932 // Ignore icmp instructions which are already being analyzed. 2933 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) { 2934 unsigned OtherIdx = !UI.getOperandNo(); 2935 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx)); 2936 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L)) 2937 continue; 2938 } 2939 2940 LSRFixup &LF = getNewFixup(); 2941 LF.UserInst = const_cast<Instruction *>(UserInst); 2942 LF.OperandValToReplace = UI.getUse(); 2943 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0); 2944 LF.LUIdx = P.first; 2945 LF.Offset = P.second; 2946 LSRUse &LU = Uses[LF.LUIdx]; 2947 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L); 2948 if (!LU.WidestFixupType || 2949 SE.getTypeSizeInBits(LU.WidestFixupType) < 2950 SE.getTypeSizeInBits(LF.OperandValToReplace->getType())) 2951 LU.WidestFixupType = LF.OperandValToReplace->getType(); 2952 InsertSupplementalFormula(U, LU, LF.LUIdx); 2953 CountRegisters(LU.Formulae.back(), Uses.size() - 1); 2954 break; 2955 } 2956 } 2957 } 2958} 2959 2960/// CollectSubexprs - Split S into subexpressions which can be pulled out into 2961/// separate registers. If C is non-null, multiply each subexpression by C. 2962static void CollectSubexprs(const SCEV *S, const SCEVConstant *C, 2963 SmallVectorImpl<const SCEV *> &Ops, 2964 const Loop *L, 2965 ScalarEvolution &SE) { 2966 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 2967 // Break out add operands. 2968 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 2969 I != E; ++I) 2970 CollectSubexprs(*I, C, Ops, L, SE); 2971 return; 2972 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 2973 // Split a non-zero base out of an addrec. 2974 if (!AR->getStart()->isZero()) { 2975 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0), 2976 AR->getStepRecurrence(SE), 2977 AR->getLoop(), 2978 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 2979 SCEV::FlagAnyWrap), 2980 C, Ops, L, SE); 2981 CollectSubexprs(AR->getStart(), C, Ops, L, SE); 2982 return; 2983 } 2984 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) { 2985 // Break (C * (a + b + c)) into C*a + C*b + C*c. 2986 if (Mul->getNumOperands() == 2) 2987 if (const SCEVConstant *Op0 = 2988 dyn_cast<SCEVConstant>(Mul->getOperand(0))) { 2989 CollectSubexprs(Mul->getOperand(1), 2990 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0, 2991 Ops, L, SE); 2992 return; 2993 } 2994 } 2995 2996 // Otherwise use the value itself, optionally with a scale applied. 2997 Ops.push_back(C ? SE.getMulExpr(C, S) : S); 2998} 2999 3000/// GenerateReassociations - Split out subexpressions from adds and the bases of 3001/// addrecs. 3002void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx, 3003 Formula Base, 3004 unsigned Depth) { 3005 // Arbitrarily cap recursion to protect compile time. 3006 if (Depth >= 3) return; 3007 3008 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { 3009 const SCEV *BaseReg = Base.BaseRegs[i]; 3010 3011 SmallVector<const SCEV *, 8> AddOps; 3012 CollectSubexprs(BaseReg, 0, AddOps, L, SE); 3013 3014 if (AddOps.size() == 1) continue; 3015 3016 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(), 3017 JE = AddOps.end(); J != JE; ++J) { 3018 3019 // Loop-variant "unknown" values are uninteresting; we won't be able to 3020 // do anything meaningful with them. 3021 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L)) 3022 continue; 3023 3024 // Don't pull a constant into a register if the constant could be folded 3025 // into an immediate field. 3026 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset, 3027 Base.getNumRegs() > 1, 3028 LU.Kind, LU.AccessTy, TLI, SE)) 3029 continue; 3030 3031 // Collect all operands except *J. 3032 SmallVector<const SCEV *, 8> InnerAddOps 3033 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J); 3034 InnerAddOps.append 3035 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end()); 3036 3037 // Don't leave just a constant behind in a register if the constant could 3038 // be folded into an immediate field. 3039 if (InnerAddOps.size() == 1 && 3040 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset, 3041 Base.getNumRegs() > 1, 3042 LU.Kind, LU.AccessTy, TLI, SE)) 3043 continue; 3044 3045 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps); 3046 if (InnerSum->isZero()) 3047 continue; 3048 Formula F = Base; 3049 3050 // Add the remaining pieces of the add back into the new formula. 3051 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum); 3052 if (TLI && InnerSumSC && 3053 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 && 3054 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset + 3055 InnerSumSC->getValue()->getZExtValue())) { 3056 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset + 3057 InnerSumSC->getValue()->getZExtValue(); 3058 F.BaseRegs.erase(F.BaseRegs.begin() + i); 3059 } else 3060 F.BaseRegs[i] = InnerSum; 3061 3062 // Add J as its own register, or an unfolded immediate. 3063 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J); 3064 if (TLI && SC && SE.getTypeSizeInBits(SC->getType()) <= 64 && 3065 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset + 3066 SC->getValue()->getZExtValue())) 3067 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset + 3068 SC->getValue()->getZExtValue(); 3069 else 3070 F.BaseRegs.push_back(*J); 3071 3072 if (InsertFormula(LU, LUIdx, F)) 3073 // If that formula hadn't been seen before, recurse to find more like 3074 // it. 3075 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1); 3076 } 3077 } 3078} 3079 3080/// GenerateCombinations - Generate a formula consisting of all of the 3081/// loop-dominating registers added into a single register. 3082void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx, 3083 Formula Base) { 3084 // This method is only interesting on a plurality of registers. 3085 if (Base.BaseRegs.size() <= 1) return; 3086 3087 Formula F = Base; 3088 F.BaseRegs.clear(); 3089 SmallVector<const SCEV *, 4> Ops; 3090 for (SmallVectorImpl<const SCEV *>::const_iterator 3091 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) { 3092 const SCEV *BaseReg = *I; 3093 if (SE.properlyDominates(BaseReg, L->getHeader()) && 3094 !SE.hasComputableLoopEvolution(BaseReg, L)) 3095 Ops.push_back(BaseReg); 3096 else 3097 F.BaseRegs.push_back(BaseReg); 3098 } 3099 if (Ops.size() > 1) { 3100 const SCEV *Sum = SE.getAddExpr(Ops); 3101 // TODO: If Sum is zero, it probably means ScalarEvolution missed an 3102 // opportunity to fold something. For now, just ignore such cases 3103 // rather than proceed with zero in a register. 3104 if (!Sum->isZero()) { 3105 F.BaseRegs.push_back(Sum); 3106 (void)InsertFormula(LU, LUIdx, F); 3107 } 3108 } 3109} 3110 3111/// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets. 3112void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, 3113 Formula Base) { 3114 // We can't add a symbolic offset if the address already contains one. 3115 if (Base.AM.BaseGV) return; 3116 3117 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { 3118 const SCEV *G = Base.BaseRegs[i]; 3119 GlobalValue *GV = ExtractSymbol(G, SE); 3120 if (G->isZero() || !GV) 3121 continue; 3122 Formula F = Base; 3123 F.AM.BaseGV = GV; 3124 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset, 3125 LU.Kind, LU.AccessTy, TLI)) 3126 continue; 3127 F.BaseRegs[i] = G; 3128 (void)InsertFormula(LU, LUIdx, F); 3129 } 3130} 3131 3132/// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets. 3133void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, 3134 Formula Base) { 3135 // TODO: For now, just add the min and max offset, because it usually isn't 3136 // worthwhile looking at everything inbetween. 3137 SmallVector<int64_t, 2> Worklist; 3138 Worklist.push_back(LU.MinOffset); 3139 if (LU.MaxOffset != LU.MinOffset) 3140 Worklist.push_back(LU.MaxOffset); 3141 3142 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { 3143 const SCEV *G = Base.BaseRegs[i]; 3144 3145 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(), 3146 E = Worklist.end(); I != E; ++I) { 3147 Formula F = Base; 3148 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I; 3149 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I, 3150 LU.Kind, LU.AccessTy, TLI)) { 3151 // Add the offset to the base register. 3152 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G); 3153 // If it cancelled out, drop the base register, otherwise update it. 3154 if (NewG->isZero()) { 3155 std::swap(F.BaseRegs[i], F.BaseRegs.back()); 3156 F.BaseRegs.pop_back(); 3157 } else 3158 F.BaseRegs[i] = NewG; 3159 3160 (void)InsertFormula(LU, LUIdx, F); 3161 } 3162 } 3163 3164 int64_t Imm = ExtractImmediate(G, SE); 3165 if (G->isZero() || Imm == 0) 3166 continue; 3167 Formula F = Base; 3168 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm; 3169 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset, 3170 LU.Kind, LU.AccessTy, TLI)) 3171 continue; 3172 F.BaseRegs[i] = G; 3173 (void)InsertFormula(LU, LUIdx, F); 3174 } 3175} 3176 3177/// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up 3178/// the comparison. For example, x == y -> x*c == y*c. 3179void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, 3180 Formula Base) { 3181 if (LU.Kind != LSRUse::ICmpZero) return; 3182 3183 // Determine the integer type for the base formula. 3184 Type *IntTy = Base.getType(); 3185 if (!IntTy) return; 3186 if (SE.getTypeSizeInBits(IntTy) > 64) return; 3187 3188 // Don't do this if there is more than one offset. 3189 if (LU.MinOffset != LU.MaxOffset) return; 3190 3191 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!"); 3192 3193 // Check each interesting stride. 3194 for (SmallSetVector<int64_t, 8>::const_iterator 3195 I = Factors.begin(), E = Factors.end(); I != E; ++I) { 3196 int64_t Factor = *I; 3197 3198 // Check that the multiplication doesn't overflow. 3199 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1) 3200 continue; 3201 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor; 3202 if (NewBaseOffs / Factor != Base.AM.BaseOffs) 3203 continue; 3204 3205 // Check that multiplying with the use offset doesn't overflow. 3206 int64_t Offset = LU.MinOffset; 3207 if (Offset == INT64_MIN && Factor == -1) 3208 continue; 3209 Offset = (uint64_t)Offset * Factor; 3210 if (Offset / Factor != LU.MinOffset) 3211 continue; 3212 3213 Formula F = Base; 3214 F.AM.BaseOffs = NewBaseOffs; 3215 3216 // Check that this scale is legal. 3217 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI)) 3218 continue; 3219 3220 // Compensate for the use having MinOffset built into it. 3221 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset; 3222 3223 const SCEV *FactorS = SE.getConstant(IntTy, Factor); 3224 3225 // Check that multiplying with each base register doesn't overflow. 3226 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) { 3227 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS); 3228 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i]) 3229 goto next; 3230 } 3231 3232 // Check that multiplying with the scaled register doesn't overflow. 3233 if (F.ScaledReg) { 3234 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS); 3235 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg) 3236 continue; 3237 } 3238 3239 // Check that multiplying with the unfolded offset doesn't overflow. 3240 if (F.UnfoldedOffset != 0) { 3241 if (F.UnfoldedOffset == INT64_MIN && Factor == -1) 3242 continue; 3243 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor; 3244 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset) 3245 continue; 3246 } 3247 3248 // If we make it here and it's legal, add it. 3249 (void)InsertFormula(LU, LUIdx, F); 3250 next:; 3251 } 3252} 3253 3254/// GenerateScales - Generate stride factor reuse formulae by making use of 3255/// scaled-offset address modes, for example. 3256void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) { 3257 // Determine the integer type for the base formula. 3258 Type *IntTy = Base.getType(); 3259 if (!IntTy) return; 3260 3261 // If this Formula already has a scaled register, we can't add another one. 3262 if (Base.AM.Scale != 0) return; 3263 3264 // Check each interesting stride. 3265 for (SmallSetVector<int64_t, 8>::const_iterator 3266 I = Factors.begin(), E = Factors.end(); I != E; ++I) { 3267 int64_t Factor = *I; 3268 3269 Base.AM.Scale = Factor; 3270 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1; 3271 // Check whether this scale is going to be legal. 3272 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset, 3273 LU.Kind, LU.AccessTy, TLI)) { 3274 // As a special-case, handle special out-of-loop Basic users specially. 3275 // TODO: Reconsider this special case. 3276 if (LU.Kind == LSRUse::Basic && 3277 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset, 3278 LSRUse::Special, LU.AccessTy, TLI) && 3279 LU.AllFixupsOutsideLoop) 3280 LU.Kind = LSRUse::Special; 3281 else 3282 continue; 3283 } 3284 // For an ICmpZero, negating a solitary base register won't lead to 3285 // new solutions. 3286 if (LU.Kind == LSRUse::ICmpZero && 3287 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV) 3288 continue; 3289 // For each addrec base reg, apply the scale, if possible. 3290 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) 3291 if (const SCEVAddRecExpr *AR = 3292 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) { 3293 const SCEV *FactorS = SE.getConstant(IntTy, Factor); 3294 if (FactorS->isZero()) 3295 continue; 3296 // Divide out the factor, ignoring high bits, since we'll be 3297 // scaling the value back up in the end. 3298 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) { 3299 // TODO: This could be optimized to avoid all the copying. 3300 Formula F = Base; 3301 F.ScaledReg = Quotient; 3302 F.DeleteBaseReg(F.BaseRegs[i]); 3303 (void)InsertFormula(LU, LUIdx, F); 3304 } 3305 } 3306 } 3307} 3308 3309/// GenerateTruncates - Generate reuse formulae from different IV types. 3310void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) { 3311 // This requires TargetLowering to tell us which truncates are free. 3312 if (!TLI) return; 3313 3314 // Don't bother truncating symbolic values. 3315 if (Base.AM.BaseGV) return; 3316 3317 // Determine the integer type for the base formula. 3318 Type *DstTy = Base.getType(); 3319 if (!DstTy) return; 3320 DstTy = SE.getEffectiveSCEVType(DstTy); 3321 3322 for (SmallSetVector<Type *, 4>::const_iterator 3323 I = Types.begin(), E = Types.end(); I != E; ++I) { 3324 Type *SrcTy = *I; 3325 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) { 3326 Formula F = Base; 3327 3328 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I); 3329 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(), 3330 JE = F.BaseRegs.end(); J != JE; ++J) 3331 *J = SE.getAnyExtendExpr(*J, SrcTy); 3332 3333 // TODO: This assumes we've done basic processing on all uses and 3334 // have an idea what the register usage is. 3335 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses)) 3336 continue; 3337 3338 (void)InsertFormula(LU, LUIdx, F); 3339 } 3340 } 3341} 3342 3343namespace { 3344 3345/// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to 3346/// defer modifications so that the search phase doesn't have to worry about 3347/// the data structures moving underneath it. 3348struct WorkItem { 3349 size_t LUIdx; 3350 int64_t Imm; 3351 const SCEV *OrigReg; 3352 3353 WorkItem(size_t LI, int64_t I, const SCEV *R) 3354 : LUIdx(LI), Imm(I), OrigReg(R) {} 3355 3356 void print(raw_ostream &OS) const; 3357 void dump() const; 3358}; 3359 3360} 3361 3362void WorkItem::print(raw_ostream &OS) const { 3363 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx 3364 << " , add offset " << Imm; 3365} 3366 3367void WorkItem::dump() const { 3368 print(errs()); errs() << '\n'; 3369} 3370 3371/// GenerateCrossUseConstantOffsets - Look for registers which are a constant 3372/// distance apart and try to form reuse opportunities between them. 3373void LSRInstance::GenerateCrossUseConstantOffsets() { 3374 // Group the registers by their value without any added constant offset. 3375 typedef std::map<int64_t, const SCEV *> ImmMapTy; 3376 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy; 3377 RegMapTy Map; 3378 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap; 3379 SmallVector<const SCEV *, 8> Sequence; 3380 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end(); 3381 I != E; ++I) { 3382 const SCEV *Reg = *I; 3383 int64_t Imm = ExtractImmediate(Reg, SE); 3384 std::pair<RegMapTy::iterator, bool> Pair = 3385 Map.insert(std::make_pair(Reg, ImmMapTy())); 3386 if (Pair.second) 3387 Sequence.push_back(Reg); 3388 Pair.first->second.insert(std::make_pair(Imm, *I)); 3389 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I); 3390 } 3391 3392 // Now examine each set of registers with the same base value. Build up 3393 // a list of work to do and do the work in a separate step so that we're 3394 // not adding formulae and register counts while we're searching. 3395 SmallVector<WorkItem, 32> WorkItems; 3396 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems; 3397 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(), 3398 E = Sequence.end(); I != E; ++I) { 3399 const SCEV *Reg = *I; 3400 const ImmMapTy &Imms = Map.find(Reg)->second; 3401 3402 // It's not worthwhile looking for reuse if there's only one offset. 3403 if (Imms.size() == 1) 3404 continue; 3405 3406 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':'; 3407 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end(); 3408 J != JE; ++J) 3409 dbgs() << ' ' << J->first; 3410 dbgs() << '\n'); 3411 3412 // Examine each offset. 3413 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end(); 3414 J != JE; ++J) { 3415 const SCEV *OrigReg = J->second; 3416 3417 int64_t JImm = J->first; 3418 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg); 3419 3420 if (!isa<SCEVConstant>(OrigReg) && 3421 UsedByIndicesMap[Reg].count() == 1) { 3422 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n'); 3423 continue; 3424 } 3425 3426 // Conservatively examine offsets between this orig reg a few selected 3427 // other orig regs. 3428 ImmMapTy::const_iterator OtherImms[] = { 3429 Imms.begin(), prior(Imms.end()), 3430 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2) 3431 }; 3432 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) { 3433 ImmMapTy::const_iterator M = OtherImms[i]; 3434 if (M == J || M == JE) continue; 3435 3436 // Compute the difference between the two. 3437 int64_t Imm = (uint64_t)JImm - M->first; 3438 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1; 3439 LUIdx = UsedByIndices.find_next(LUIdx)) 3440 // Make a memo of this use, offset, and register tuple. 3441 if (UniqueItems.insert(std::make_pair(LUIdx, Imm))) 3442 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg)); 3443 } 3444 } 3445 } 3446 3447 Map.clear(); 3448 Sequence.clear(); 3449 UsedByIndicesMap.clear(); 3450 UniqueItems.clear(); 3451 3452 // Now iterate through the worklist and add new formulae. 3453 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(), 3454 E = WorkItems.end(); I != E; ++I) { 3455 const WorkItem &WI = *I; 3456 size_t LUIdx = WI.LUIdx; 3457 LSRUse &LU = Uses[LUIdx]; 3458 int64_t Imm = WI.Imm; 3459 const SCEV *OrigReg = WI.OrigReg; 3460 3461 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType()); 3462 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm)); 3463 unsigned BitWidth = SE.getTypeSizeInBits(IntTy); 3464 3465 // TODO: Use a more targeted data structure. 3466 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) { 3467 const Formula &F = LU.Formulae[L]; 3468 // Use the immediate in the scaled register. 3469 if (F.ScaledReg == OrigReg) { 3470 int64_t Offs = (uint64_t)F.AM.BaseOffs + 3471 Imm * (uint64_t)F.AM.Scale; 3472 // Don't create 50 + reg(-50). 3473 if (F.referencesReg(SE.getSCEV( 3474 ConstantInt::get(IntTy, -(uint64_t)Offs)))) 3475 continue; 3476 Formula NewF = F; 3477 NewF.AM.BaseOffs = Offs; 3478 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset, 3479 LU.Kind, LU.AccessTy, TLI)) 3480 continue; 3481 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg); 3482 3483 // If the new scale is a constant in a register, and adding the constant 3484 // value to the immediate would produce a value closer to zero than the 3485 // immediate itself, then the formula isn't worthwhile. 3486 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg)) 3487 if (C->getValue()->isNegative() != 3488 (NewF.AM.BaseOffs < 0) && 3489 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale)) 3490 .ule(abs64(NewF.AM.BaseOffs))) 3491 continue; 3492 3493 // OK, looks good. 3494 (void)InsertFormula(LU, LUIdx, NewF); 3495 } else { 3496 // Use the immediate in a base register. 3497 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) { 3498 const SCEV *BaseReg = F.BaseRegs[N]; 3499 if (BaseReg != OrigReg) 3500 continue; 3501 Formula NewF = F; 3502 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm; 3503 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset, 3504 LU.Kind, LU.AccessTy, TLI)) { 3505 if (!TLI || 3506 !TLI->isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm)) 3507 continue; 3508 NewF = F; 3509 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm; 3510 } 3511 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg); 3512 3513 // If the new formula has a constant in a register, and adding the 3514 // constant value to the immediate would produce a value closer to 3515 // zero than the immediate itself, then the formula isn't worthwhile. 3516 for (SmallVectorImpl<const SCEV *>::const_iterator 3517 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end(); 3518 J != JE; ++J) 3519 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J)) 3520 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt( 3521 abs64(NewF.AM.BaseOffs)) && 3522 (C->getValue()->getValue() + 3523 NewF.AM.BaseOffs).countTrailingZeros() >= 3524 CountTrailingZeros_64(NewF.AM.BaseOffs)) 3525 goto skip_formula; 3526 3527 // Ok, looks good. 3528 (void)InsertFormula(LU, LUIdx, NewF); 3529 break; 3530 skip_formula:; 3531 } 3532 } 3533 } 3534 } 3535} 3536 3537/// GenerateAllReuseFormulae - Generate formulae for each use. 3538void 3539LSRInstance::GenerateAllReuseFormulae() { 3540 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan 3541 // queries are more precise. 3542 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3543 LSRUse &LU = Uses[LUIdx]; 3544 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3545 GenerateReassociations(LU, LUIdx, LU.Formulae[i]); 3546 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3547 GenerateCombinations(LU, LUIdx, LU.Formulae[i]); 3548 } 3549 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3550 LSRUse &LU = Uses[LUIdx]; 3551 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3552 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]); 3553 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3554 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]); 3555 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3556 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]); 3557 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3558 GenerateScales(LU, LUIdx, LU.Formulae[i]); 3559 } 3560 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3561 LSRUse &LU = Uses[LUIdx]; 3562 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3563 GenerateTruncates(LU, LUIdx, LU.Formulae[i]); 3564 } 3565 3566 GenerateCrossUseConstantOffsets(); 3567 3568 DEBUG(dbgs() << "\n" 3569 "After generating reuse formulae:\n"; 3570 print_uses(dbgs())); 3571} 3572 3573/// If there are multiple formulae with the same set of registers used 3574/// by other uses, pick the best one and delete the others. 3575void LSRInstance::FilterOutUndesirableDedicatedRegisters() { 3576 DenseSet<const SCEV *> VisitedRegs; 3577 SmallPtrSet<const SCEV *, 16> Regs; 3578 SmallPtrSet<const SCEV *, 16> LoserRegs; 3579#ifndef NDEBUG 3580 bool ChangedFormulae = false; 3581#endif 3582 3583 // Collect the best formula for each unique set of shared registers. This 3584 // is reset for each use. 3585 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo> 3586 BestFormulaeTy; 3587 BestFormulaeTy BestFormulae; 3588 3589 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3590 LSRUse &LU = Uses[LUIdx]; 3591 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n'); 3592 3593 bool Any = false; 3594 for (size_t FIdx = 0, NumForms = LU.Formulae.size(); 3595 FIdx != NumForms; ++FIdx) { 3596 Formula &F = LU.Formulae[FIdx]; 3597 3598 // Some formulas are instant losers. For example, they may depend on 3599 // nonexistent AddRecs from other loops. These need to be filtered 3600 // immediately, otherwise heuristics could choose them over others leading 3601 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here 3602 // avoids the need to recompute this information across formulae using the 3603 // same bad AddRec. Passing LoserRegs is also essential unless we remove 3604 // the corresponding bad register from the Regs set. 3605 Cost CostF; 3606 Regs.clear(); 3607 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT, 3608 &LoserRegs); 3609 if (CostF.isLoser()) { 3610 // During initial formula generation, undesirable formulae are generated 3611 // by uses within other loops that have some non-trivial address mode or 3612 // use the postinc form of the IV. LSR needs to provide these formulae 3613 // as the basis of rediscovering the desired formula that uses an AddRec 3614 // corresponding to the existing phi. Once all formulae have been 3615 // generated, these initial losers may be pruned. 3616 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs()); 3617 dbgs() << "\n"); 3618 } 3619 else { 3620 SmallVector<const SCEV *, 2> Key; 3621 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(), 3622 JE = F.BaseRegs.end(); J != JE; ++J) { 3623 const SCEV *Reg = *J; 3624 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx)) 3625 Key.push_back(Reg); 3626 } 3627 if (F.ScaledReg && 3628 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx)) 3629 Key.push_back(F.ScaledReg); 3630 // Unstable sort by host order ok, because this is only used for 3631 // uniquifying. 3632 std::sort(Key.begin(), Key.end()); 3633 3634 std::pair<BestFormulaeTy::const_iterator, bool> P = 3635 BestFormulae.insert(std::make_pair(Key, FIdx)); 3636 if (P.second) 3637 continue; 3638 3639 Formula &Best = LU.Formulae[P.first->second]; 3640 3641 Cost CostBest; 3642 Regs.clear(); 3643 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT); 3644 if (CostF < CostBest) 3645 std::swap(F, Best); 3646 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs()); 3647 dbgs() << "\n" 3648 " in favor of formula "; Best.print(dbgs()); 3649 dbgs() << '\n'); 3650 } 3651#ifndef NDEBUG 3652 ChangedFormulae = true; 3653#endif 3654 LU.DeleteFormula(F); 3655 --FIdx; 3656 --NumForms; 3657 Any = true; 3658 } 3659 3660 // Now that we've filtered out some formulae, recompute the Regs set. 3661 if (Any) 3662 LU.RecomputeRegs(LUIdx, RegUses); 3663 3664 // Reset this to prepare for the next use. 3665 BestFormulae.clear(); 3666 } 3667 3668 DEBUG(if (ChangedFormulae) { 3669 dbgs() << "\n" 3670 "After filtering out undesirable candidates:\n"; 3671 print_uses(dbgs()); 3672 }); 3673} 3674 3675// This is a rough guess that seems to work fairly well. 3676static const size_t ComplexityLimit = UINT16_MAX; 3677 3678/// EstimateSearchSpaceComplexity - Estimate the worst-case number of 3679/// solutions the solver might have to consider. It almost never considers 3680/// this many solutions because it prune the search space, but the pruning 3681/// isn't always sufficient. 3682size_t LSRInstance::EstimateSearchSpaceComplexity() const { 3683 size_t Power = 1; 3684 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), 3685 E = Uses.end(); I != E; ++I) { 3686 size_t FSize = I->Formulae.size(); 3687 if (FSize >= ComplexityLimit) { 3688 Power = ComplexityLimit; 3689 break; 3690 } 3691 Power *= FSize; 3692 if (Power >= ComplexityLimit) 3693 break; 3694 } 3695 return Power; 3696} 3697 3698/// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset 3699/// of the registers of another formula, it won't help reduce register 3700/// pressure (though it may not necessarily hurt register pressure); remove 3701/// it to simplify the system. 3702void LSRInstance::NarrowSearchSpaceByDetectingSupersets() { 3703 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3704 DEBUG(dbgs() << "The search space is too complex.\n"); 3705 3706 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae " 3707 "which use a superset of registers used by other " 3708 "formulae.\n"); 3709 3710 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3711 LSRUse &LU = Uses[LUIdx]; 3712 bool Any = false; 3713 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) { 3714 Formula &F = LU.Formulae[i]; 3715 // Look for a formula with a constant or GV in a register. If the use 3716 // also has a formula with that same value in an immediate field, 3717 // delete the one that uses a register. 3718 for (SmallVectorImpl<const SCEV *>::const_iterator 3719 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) { 3720 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) { 3721 Formula NewF = F; 3722 NewF.AM.BaseOffs += C->getValue()->getSExtValue(); 3723 NewF.BaseRegs.erase(NewF.BaseRegs.begin() + 3724 (I - F.BaseRegs.begin())); 3725 if (LU.HasFormulaWithSameRegs(NewF)) { 3726 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n'); 3727 LU.DeleteFormula(F); 3728 --i; 3729 --e; 3730 Any = true; 3731 break; 3732 } 3733 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) { 3734 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) 3735 if (!F.AM.BaseGV) { 3736 Formula NewF = F; 3737 NewF.AM.BaseGV = GV; 3738 NewF.BaseRegs.erase(NewF.BaseRegs.begin() + 3739 (I - F.BaseRegs.begin())); 3740 if (LU.HasFormulaWithSameRegs(NewF)) { 3741 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); 3742 dbgs() << '\n'); 3743 LU.DeleteFormula(F); 3744 --i; 3745 --e; 3746 Any = true; 3747 break; 3748 } 3749 } 3750 } 3751 } 3752 } 3753 if (Any) 3754 LU.RecomputeRegs(LUIdx, RegUses); 3755 } 3756 3757 DEBUG(dbgs() << "After pre-selection:\n"; 3758 print_uses(dbgs())); 3759 } 3760} 3761 3762/// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers 3763/// for expressions like A, A+1, A+2, etc., allocate a single register for 3764/// them. 3765void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() { 3766 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3767 DEBUG(dbgs() << "The search space is too complex.\n"); 3768 3769 DEBUG(dbgs() << "Narrowing the search space by assuming that uses " 3770 "separated by a constant offset will use the same " 3771 "registers.\n"); 3772 3773 // This is especially useful for unrolled loops. 3774 3775 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3776 LSRUse &LU = Uses[LUIdx]; 3777 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(), 3778 E = LU.Formulae.end(); I != E; ++I) { 3779 const Formula &F = *I; 3780 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) { 3781 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) { 3782 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs, 3783 /*HasBaseReg=*/false, 3784 LU.Kind, LU.AccessTy)) { 3785 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); 3786 dbgs() << '\n'); 3787 3788 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop; 3789 3790 // Update the relocs to reference the new use. 3791 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(), 3792 E = Fixups.end(); I != E; ++I) { 3793 LSRFixup &Fixup = *I; 3794 if (Fixup.LUIdx == LUIdx) { 3795 Fixup.LUIdx = LUThatHas - &Uses.front(); 3796 Fixup.Offset += F.AM.BaseOffs; 3797 // Add the new offset to LUThatHas' offset list. 3798 if (LUThatHas->Offsets.back() != Fixup.Offset) { 3799 LUThatHas->Offsets.push_back(Fixup.Offset); 3800 if (Fixup.Offset > LUThatHas->MaxOffset) 3801 LUThatHas->MaxOffset = Fixup.Offset; 3802 if (Fixup.Offset < LUThatHas->MinOffset) 3803 LUThatHas->MinOffset = Fixup.Offset; 3804 } 3805 DEBUG(dbgs() << "New fixup has offset " 3806 << Fixup.Offset << '\n'); 3807 } 3808 if (Fixup.LUIdx == NumUses-1) 3809 Fixup.LUIdx = LUIdx; 3810 } 3811 3812 // Delete formulae from the new use which are no longer legal. 3813 bool Any = false; 3814 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) { 3815 Formula &F = LUThatHas->Formulae[i]; 3816 if (!isLegalUse(F.AM, 3817 LUThatHas->MinOffset, LUThatHas->MaxOffset, 3818 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) { 3819 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); 3820 dbgs() << '\n'); 3821 LUThatHas->DeleteFormula(F); 3822 --i; 3823 --e; 3824 Any = true; 3825 } 3826 } 3827 if (Any) 3828 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses); 3829 3830 // Delete the old use. 3831 DeleteUse(LU, LUIdx); 3832 --LUIdx; 3833 --NumUses; 3834 break; 3835 } 3836 } 3837 } 3838 } 3839 } 3840 3841 DEBUG(dbgs() << "After pre-selection:\n"; 3842 print_uses(dbgs())); 3843 } 3844} 3845 3846/// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call 3847/// FilterOutUndesirableDedicatedRegisters again, if necessary, now that 3848/// we've done more filtering, as it may be able to find more formulae to 3849/// eliminate. 3850void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){ 3851 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3852 DEBUG(dbgs() << "The search space is too complex.\n"); 3853 3854 DEBUG(dbgs() << "Narrowing the search space by re-filtering out " 3855 "undesirable dedicated registers.\n"); 3856 3857 FilterOutUndesirableDedicatedRegisters(); 3858 3859 DEBUG(dbgs() << "After pre-selection:\n"; 3860 print_uses(dbgs())); 3861 } 3862} 3863 3864/// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely 3865/// to be profitable, and then in any use which has any reference to that 3866/// register, delete all formulae which do not reference that register. 3867void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() { 3868 // With all other options exhausted, loop until the system is simple 3869 // enough to handle. 3870 SmallPtrSet<const SCEV *, 4> Taken; 3871 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3872 // Ok, we have too many of formulae on our hands to conveniently handle. 3873 // Use a rough heuristic to thin out the list. 3874 DEBUG(dbgs() << "The search space is too complex.\n"); 3875 3876 // Pick the register which is used by the most LSRUses, which is likely 3877 // to be a good reuse register candidate. 3878 const SCEV *Best = 0; 3879 unsigned BestNum = 0; 3880 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end(); 3881 I != E; ++I) { 3882 const SCEV *Reg = *I; 3883 if (Taken.count(Reg)) 3884 continue; 3885 if (!Best) 3886 Best = Reg; 3887 else { 3888 unsigned Count = RegUses.getUsedByIndices(Reg).count(); 3889 if (Count > BestNum) { 3890 Best = Reg; 3891 BestNum = Count; 3892 } 3893 } 3894 } 3895 3896 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best 3897 << " will yield profitable reuse.\n"); 3898 Taken.insert(Best); 3899 3900 // In any use with formulae which references this register, delete formulae 3901 // which don't reference it. 3902 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3903 LSRUse &LU = Uses[LUIdx]; 3904 if (!LU.Regs.count(Best)) continue; 3905 3906 bool Any = false; 3907 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) { 3908 Formula &F = LU.Formulae[i]; 3909 if (!F.referencesReg(Best)) { 3910 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n'); 3911 LU.DeleteFormula(F); 3912 --e; 3913 --i; 3914 Any = true; 3915 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?"); 3916 continue; 3917 } 3918 } 3919 3920 if (Any) 3921 LU.RecomputeRegs(LUIdx, RegUses); 3922 } 3923 3924 DEBUG(dbgs() << "After pre-selection:\n"; 3925 print_uses(dbgs())); 3926 } 3927} 3928 3929/// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of 3930/// formulae to choose from, use some rough heuristics to prune down the number 3931/// of formulae. This keeps the main solver from taking an extraordinary amount 3932/// of time in some worst-case scenarios. 3933void LSRInstance::NarrowSearchSpaceUsingHeuristics() { 3934 NarrowSearchSpaceByDetectingSupersets(); 3935 NarrowSearchSpaceByCollapsingUnrolledCode(); 3936 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(); 3937 NarrowSearchSpaceByPickingWinnerRegs(); 3938} 3939 3940/// SolveRecurse - This is the recursive solver. 3941void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution, 3942 Cost &SolutionCost, 3943 SmallVectorImpl<const Formula *> &Workspace, 3944 const Cost &CurCost, 3945 const SmallPtrSet<const SCEV *, 16> &CurRegs, 3946 DenseSet<const SCEV *> &VisitedRegs) const { 3947 // Some ideas: 3948 // - prune more: 3949 // - use more aggressive filtering 3950 // - sort the formula so that the most profitable solutions are found first 3951 // - sort the uses too 3952 // - search faster: 3953 // - don't compute a cost, and then compare. compare while computing a cost 3954 // and bail early. 3955 // - track register sets with SmallBitVector 3956 3957 const LSRUse &LU = Uses[Workspace.size()]; 3958 3959 // If this use references any register that's already a part of the 3960 // in-progress solution, consider it a requirement that a formula must 3961 // reference that register in order to be considered. This prunes out 3962 // unprofitable searching. 3963 SmallSetVector<const SCEV *, 4> ReqRegs; 3964 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(), 3965 E = CurRegs.end(); I != E; ++I) 3966 if (LU.Regs.count(*I)) 3967 ReqRegs.insert(*I); 3968 3969 SmallPtrSet<const SCEV *, 16> NewRegs; 3970 Cost NewCost; 3971 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(), 3972 E = LU.Formulae.end(); I != E; ++I) { 3973 const Formula &F = *I; 3974 3975 // Ignore formulae which do not use any of the required registers. 3976 bool SatisfiedReqReg = true; 3977 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(), 3978 JE = ReqRegs.end(); J != JE; ++J) { 3979 const SCEV *Reg = *J; 3980 if ((!F.ScaledReg || F.ScaledReg != Reg) && 3981 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) == 3982 F.BaseRegs.end()) { 3983 SatisfiedReqReg = false; 3984 break; 3985 } 3986 } 3987 if (!SatisfiedReqReg) { 3988 // If none of the formulae satisfied the required registers, then we could 3989 // clear ReqRegs and try again. Currently, we simply give up in this case. 3990 continue; 3991 } 3992 3993 // Evaluate the cost of the current formula. If it's already worse than 3994 // the current best, prune the search at that point. 3995 NewCost = CurCost; 3996 NewRegs = CurRegs; 3997 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT); 3998 if (NewCost < SolutionCost) { 3999 Workspace.push_back(&F); 4000 if (Workspace.size() != Uses.size()) { 4001 SolveRecurse(Solution, SolutionCost, Workspace, NewCost, 4002 NewRegs, VisitedRegs); 4003 if (F.getNumRegs() == 1 && Workspace.size() == 1) 4004 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]); 4005 } else { 4006 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs()); 4007 dbgs() << ".\n Regs:"; 4008 for (SmallPtrSet<const SCEV *, 16>::const_iterator 4009 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I) 4010 dbgs() << ' ' << **I; 4011 dbgs() << '\n'); 4012 4013 SolutionCost = NewCost; 4014 Solution = Workspace; 4015 } 4016 Workspace.pop_back(); 4017 } 4018 } 4019} 4020 4021/// Solve - Choose one formula from each use. Return the results in the given 4022/// Solution vector. 4023void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const { 4024 SmallVector<const Formula *, 8> Workspace; 4025 Cost SolutionCost; 4026 SolutionCost.Loose(); 4027 Cost CurCost; 4028 SmallPtrSet<const SCEV *, 16> CurRegs; 4029 DenseSet<const SCEV *> VisitedRegs; 4030 Workspace.reserve(Uses.size()); 4031 4032 // SolveRecurse does all the work. 4033 SolveRecurse(Solution, SolutionCost, Workspace, CurCost, 4034 CurRegs, VisitedRegs); 4035 if (Solution.empty()) { 4036 DEBUG(dbgs() << "\nNo Satisfactory Solution\n"); 4037 return; 4038 } 4039 4040 // Ok, we've now made all our decisions. 4041 DEBUG(dbgs() << "\n" 4042 "The chosen solution requires "; SolutionCost.print(dbgs()); 4043 dbgs() << ":\n"; 4044 for (size_t i = 0, e = Uses.size(); i != e; ++i) { 4045 dbgs() << " "; 4046 Uses[i].print(dbgs()); 4047 dbgs() << "\n" 4048 " "; 4049 Solution[i]->print(dbgs()); 4050 dbgs() << '\n'; 4051 }); 4052 4053 assert(Solution.size() == Uses.size() && "Malformed solution!"); 4054} 4055 4056/// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up 4057/// the dominator tree far as we can go while still being dominated by the 4058/// input positions. This helps canonicalize the insert position, which 4059/// encourages sharing. 4060BasicBlock::iterator 4061LSRInstance::HoistInsertPosition(BasicBlock::iterator IP, 4062 const SmallVectorImpl<Instruction *> &Inputs) 4063 const { 4064 for (;;) { 4065 const Loop *IPLoop = LI.getLoopFor(IP->getParent()); 4066 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0; 4067 4068 BasicBlock *IDom; 4069 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) { 4070 if (!Rung) return IP; 4071 Rung = Rung->getIDom(); 4072 if (!Rung) return IP; 4073 IDom = Rung->getBlock(); 4074 4075 // Don't climb into a loop though. 4076 const Loop *IDomLoop = LI.getLoopFor(IDom); 4077 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0; 4078 if (IDomDepth <= IPLoopDepth && 4079 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop)) 4080 break; 4081 } 4082 4083 bool AllDominate = true; 4084 Instruction *BetterPos = 0; 4085 Instruction *Tentative = IDom->getTerminator(); 4086 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(), 4087 E = Inputs.end(); I != E; ++I) { 4088 Instruction *Inst = *I; 4089 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) { 4090 AllDominate = false; 4091 break; 4092 } 4093 // Attempt to find an insert position in the middle of the block, 4094 // instead of at the end, so that it can be used for other expansions. 4095 if (IDom == Inst->getParent() && 4096 (!BetterPos || DT.dominates(BetterPos, Inst))) 4097 BetterPos = llvm::next(BasicBlock::iterator(Inst)); 4098 } 4099 if (!AllDominate) 4100 break; 4101 if (BetterPos) 4102 IP = BetterPos; 4103 else 4104 IP = Tentative; 4105 } 4106 4107 return IP; 4108} 4109 4110/// AdjustInsertPositionForExpand - Determine an input position which will be 4111/// dominated by the operands and which will dominate the result. 4112BasicBlock::iterator 4113LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP, 4114 const LSRFixup &LF, 4115 const LSRUse &LU, 4116 SCEVExpander &Rewriter) const { 4117 // Collect some instructions which must be dominated by the 4118 // expanding replacement. These must be dominated by any operands that 4119 // will be required in the expansion. 4120 SmallVector<Instruction *, 4> Inputs; 4121 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace)) 4122 Inputs.push_back(I); 4123 if (LU.Kind == LSRUse::ICmpZero) 4124 if (Instruction *I = 4125 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1))) 4126 Inputs.push_back(I); 4127 if (LF.PostIncLoops.count(L)) { 4128 if (LF.isUseFullyOutsideLoop(L)) 4129 Inputs.push_back(L->getLoopLatch()->getTerminator()); 4130 else 4131 Inputs.push_back(IVIncInsertPos); 4132 } 4133 // The expansion must also be dominated by the increment positions of any 4134 // loops it for which it is using post-inc mode. 4135 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(), 4136 E = LF.PostIncLoops.end(); I != E; ++I) { 4137 const Loop *PIL = *I; 4138 if (PIL == L) continue; 4139 4140 // Be dominated by the loop exit. 4141 SmallVector<BasicBlock *, 4> ExitingBlocks; 4142 PIL->getExitingBlocks(ExitingBlocks); 4143 if (!ExitingBlocks.empty()) { 4144 BasicBlock *BB = ExitingBlocks[0]; 4145 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i) 4146 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]); 4147 Inputs.push_back(BB->getTerminator()); 4148 } 4149 } 4150 4151 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP) 4152 && !isa<DbgInfoIntrinsic>(LowestIP) && 4153 "Insertion point must be a normal instruction"); 4154 4155 // Then, climb up the immediate dominator tree as far as we can go while 4156 // still being dominated by the input positions. 4157 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs); 4158 4159 // Don't insert instructions before PHI nodes. 4160 while (isa<PHINode>(IP)) ++IP; 4161 4162 // Ignore landingpad instructions. 4163 while (isa<LandingPadInst>(IP)) ++IP; 4164 4165 // Ignore debug intrinsics. 4166 while (isa<DbgInfoIntrinsic>(IP)) ++IP; 4167 4168 // Set IP below instructions recently inserted by SCEVExpander. This keeps the 4169 // IP consistent across expansions and allows the previously inserted 4170 // instructions to be reused by subsequent expansion. 4171 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP; 4172 4173 return IP; 4174} 4175 4176/// Expand - Emit instructions for the leading candidate expression for this 4177/// LSRUse (this is called "expanding"). 4178Value *LSRInstance::Expand(const LSRFixup &LF, 4179 const Formula &F, 4180 BasicBlock::iterator IP, 4181 SCEVExpander &Rewriter, 4182 SmallVectorImpl<WeakVH> &DeadInsts) const { 4183 const LSRUse &LU = Uses[LF.LUIdx]; 4184 4185 // Determine an input position which will be dominated by the operands and 4186 // which will dominate the result. 4187 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter); 4188 4189 // Inform the Rewriter if we have a post-increment use, so that it can 4190 // perform an advantageous expansion. 4191 Rewriter.setPostInc(LF.PostIncLoops); 4192 4193 // This is the type that the user actually needs. 4194 Type *OpTy = LF.OperandValToReplace->getType(); 4195 // This will be the type that we'll initially expand to. 4196 Type *Ty = F.getType(); 4197 if (!Ty) 4198 // No type known; just expand directly to the ultimate type. 4199 Ty = OpTy; 4200 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy)) 4201 // Expand directly to the ultimate type if it's the right size. 4202 Ty = OpTy; 4203 // This is the type to do integer arithmetic in. 4204 Type *IntTy = SE.getEffectiveSCEVType(Ty); 4205 4206 // Build up a list of operands to add together to form the full base. 4207 SmallVector<const SCEV *, 8> Ops; 4208 4209 // Expand the BaseRegs portion. 4210 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(), 4211 E = F.BaseRegs.end(); I != E; ++I) { 4212 const SCEV *Reg = *I; 4213 assert(!Reg->isZero() && "Zero allocated in a base register!"); 4214 4215 // If we're expanding for a post-inc user, make the post-inc adjustment. 4216 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops); 4217 Reg = TransformForPostIncUse(Denormalize, Reg, 4218 LF.UserInst, LF.OperandValToReplace, 4219 Loops, SE, DT); 4220 4221 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP))); 4222 } 4223 4224 // Flush the operand list to suppress SCEVExpander hoisting. 4225 if (!Ops.empty()) { 4226 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP); 4227 Ops.clear(); 4228 Ops.push_back(SE.getUnknown(FullV)); 4229 } 4230 4231 // Expand the ScaledReg portion. 4232 Value *ICmpScaledV = 0; 4233 if (F.AM.Scale != 0) { 4234 const SCEV *ScaledS = F.ScaledReg; 4235 4236 // If we're expanding for a post-inc user, make the post-inc adjustment. 4237 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops); 4238 ScaledS = TransformForPostIncUse(Denormalize, ScaledS, 4239 LF.UserInst, LF.OperandValToReplace, 4240 Loops, SE, DT); 4241 4242 if (LU.Kind == LSRUse::ICmpZero) { 4243 // An interesting way of "folding" with an icmp is to use a negated 4244 // scale, which we'll implement by inserting it into the other operand 4245 // of the icmp. 4246 assert(F.AM.Scale == -1 && 4247 "The only scale supported by ICmpZero uses is -1!"); 4248 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP); 4249 } else { 4250 // Otherwise just expand the scaled register and an explicit scale, 4251 // which is expected to be matched as part of the address. 4252 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP)); 4253 ScaledS = SE.getMulExpr(ScaledS, 4254 SE.getConstant(ScaledS->getType(), F.AM.Scale)); 4255 Ops.push_back(ScaledS); 4256 4257 // Flush the operand list to suppress SCEVExpander hoisting. 4258 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP); 4259 Ops.clear(); 4260 Ops.push_back(SE.getUnknown(FullV)); 4261 } 4262 } 4263 4264 // Expand the GV portion. 4265 if (F.AM.BaseGV) { 4266 Ops.push_back(SE.getUnknown(F.AM.BaseGV)); 4267 4268 // Flush the operand list to suppress SCEVExpander hoisting. 4269 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP); 4270 Ops.clear(); 4271 Ops.push_back(SE.getUnknown(FullV)); 4272 } 4273 4274 // Expand the immediate portion. 4275 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset; 4276 if (Offset != 0) { 4277 if (LU.Kind == LSRUse::ICmpZero) { 4278 // The other interesting way of "folding" with an ICmpZero is to use a 4279 // negated immediate. 4280 if (!ICmpScaledV) 4281 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset); 4282 else { 4283 Ops.push_back(SE.getUnknown(ICmpScaledV)); 4284 ICmpScaledV = ConstantInt::get(IntTy, Offset); 4285 } 4286 } else { 4287 // Just add the immediate values. These again are expected to be matched 4288 // as part of the address. 4289 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset))); 4290 } 4291 } 4292 4293 // Expand the unfolded offset portion. 4294 int64_t UnfoldedOffset = F.UnfoldedOffset; 4295 if (UnfoldedOffset != 0) { 4296 // Just add the immediate values. 4297 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, 4298 UnfoldedOffset))); 4299 } 4300 4301 // Emit instructions summing all the operands. 4302 const SCEV *FullS = Ops.empty() ? 4303 SE.getConstant(IntTy, 0) : 4304 SE.getAddExpr(Ops); 4305 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP); 4306 4307 // We're done expanding now, so reset the rewriter. 4308 Rewriter.clearPostInc(); 4309 4310 // An ICmpZero Formula represents an ICmp which we're handling as a 4311 // comparison against zero. Now that we've expanded an expression for that 4312 // form, update the ICmp's other operand. 4313 if (LU.Kind == LSRUse::ICmpZero) { 4314 ICmpInst *CI = cast<ICmpInst>(LF.UserInst); 4315 DeadInsts.push_back(CI->getOperand(1)); 4316 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and " 4317 "a scale at the same time!"); 4318 if (F.AM.Scale == -1) { 4319 if (ICmpScaledV->getType() != OpTy) { 4320 Instruction *Cast = 4321 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false, 4322 OpTy, false), 4323 ICmpScaledV, OpTy, "tmp", CI); 4324 ICmpScaledV = Cast; 4325 } 4326 CI->setOperand(1, ICmpScaledV); 4327 } else { 4328 assert(F.AM.Scale == 0 && 4329 "ICmp does not support folding a global value and " 4330 "a scale at the same time!"); 4331 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy), 4332 -(uint64_t)Offset); 4333 if (C->getType() != OpTy) 4334 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 4335 OpTy, false), 4336 C, OpTy); 4337 4338 CI->setOperand(1, C); 4339 } 4340 } 4341 4342 return FullV; 4343} 4344 4345/// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use 4346/// of their operands effectively happens in their predecessor blocks, so the 4347/// expression may need to be expanded in multiple places. 4348void LSRInstance::RewriteForPHI(PHINode *PN, 4349 const LSRFixup &LF, 4350 const Formula &F, 4351 SCEVExpander &Rewriter, 4352 SmallVectorImpl<WeakVH> &DeadInsts, 4353 Pass *P) const { 4354 DenseMap<BasicBlock *, Value *> Inserted; 4355 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 4356 if (PN->getIncomingValue(i) == LF.OperandValToReplace) { 4357 BasicBlock *BB = PN->getIncomingBlock(i); 4358 4359 // If this is a critical edge, split the edge so that we do not insert 4360 // the code on all predecessor/successor paths. We do this unless this 4361 // is the canonical backedge for this loop, which complicates post-inc 4362 // users. 4363 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 && 4364 !isa<IndirectBrInst>(BB->getTerminator())) { 4365 BasicBlock *Parent = PN->getParent(); 4366 Loop *PNLoop = LI.getLoopFor(Parent); 4367 if (!PNLoop || Parent != PNLoop->getHeader()) { 4368 // Split the critical edge. 4369 BasicBlock *NewBB = 0; 4370 if (!Parent->isLandingPad()) { 4371 NewBB = SplitCriticalEdge(BB, Parent, P, 4372 /*MergeIdenticalEdges=*/true, 4373 /*DontDeleteUselessPhis=*/true); 4374 } else { 4375 SmallVector<BasicBlock*, 2> NewBBs; 4376 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs); 4377 NewBB = NewBBs[0]; 4378 } 4379 4380 // If PN is outside of the loop and BB is in the loop, we want to 4381 // move the block to be immediately before the PHI block, not 4382 // immediately after BB. 4383 if (L->contains(BB) && !L->contains(PN)) 4384 NewBB->moveBefore(PN->getParent()); 4385 4386 // Splitting the edge can reduce the number of PHI entries we have. 4387 e = PN->getNumIncomingValues(); 4388 BB = NewBB; 4389 i = PN->getBasicBlockIndex(BB); 4390 } 4391 } 4392 4393 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair = 4394 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0))); 4395 if (!Pair.second) 4396 PN->setIncomingValue(i, Pair.first->second); 4397 else { 4398 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts); 4399 4400 // If this is reuse-by-noop-cast, insert the noop cast. 4401 Type *OpTy = LF.OperandValToReplace->getType(); 4402 if (FullV->getType() != OpTy) 4403 FullV = 4404 CastInst::Create(CastInst::getCastOpcode(FullV, false, 4405 OpTy, false), 4406 FullV, LF.OperandValToReplace->getType(), 4407 "tmp", BB->getTerminator()); 4408 4409 PN->setIncomingValue(i, FullV); 4410 Pair.first->second = FullV; 4411 } 4412 } 4413} 4414 4415/// Rewrite - Emit instructions for the leading candidate expression for this 4416/// LSRUse (this is called "expanding"), and update the UserInst to reference 4417/// the newly expanded value. 4418void LSRInstance::Rewrite(const LSRFixup &LF, 4419 const Formula &F, 4420 SCEVExpander &Rewriter, 4421 SmallVectorImpl<WeakVH> &DeadInsts, 4422 Pass *P) const { 4423 // First, find an insertion point that dominates UserInst. For PHI nodes, 4424 // find the nearest block which dominates all the relevant uses. 4425 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) { 4426 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P); 4427 } else { 4428 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts); 4429 4430 // If this is reuse-by-noop-cast, insert the noop cast. 4431 Type *OpTy = LF.OperandValToReplace->getType(); 4432 if (FullV->getType() != OpTy) { 4433 Instruction *Cast = 4434 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false), 4435 FullV, OpTy, "tmp", LF.UserInst); 4436 FullV = Cast; 4437 } 4438 4439 // Update the user. ICmpZero is handled specially here (for now) because 4440 // Expand may have updated one of the operands of the icmp already, and 4441 // its new value may happen to be equal to LF.OperandValToReplace, in 4442 // which case doing replaceUsesOfWith leads to replacing both operands 4443 // with the same value. TODO: Reorganize this. 4444 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero) 4445 LF.UserInst->setOperand(0, FullV); 4446 else 4447 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV); 4448 } 4449 4450 DeadInsts.push_back(LF.OperandValToReplace); 4451} 4452 4453/// ImplementSolution - Rewrite all the fixup locations with new values, 4454/// following the chosen solution. 4455void 4456LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution, 4457 Pass *P) { 4458 // Keep track of instructions we may have made dead, so that 4459 // we can remove them after we are done working. 4460 SmallVector<WeakVH, 16> DeadInsts; 4461 4462 SCEVExpander Rewriter(SE, "lsr"); 4463#ifndef NDEBUG 4464 Rewriter.setDebugType(DEBUG_TYPE); 4465#endif 4466 Rewriter.disableCanonicalMode(); 4467 Rewriter.enableLSRMode(); 4468 Rewriter.setIVIncInsertPos(L, IVIncInsertPos); 4469 4470 // Mark phi nodes that terminate chains so the expander tries to reuse them. 4471 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(), 4472 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) { 4473 if (PHINode *PN = dyn_cast<PHINode>(ChainI->back().UserInst)) 4474 Rewriter.setChainedPhi(PN); 4475 } 4476 4477 // Expand the new value definitions and update the users. 4478 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(), 4479 E = Fixups.end(); I != E; ++I) { 4480 const LSRFixup &Fixup = *I; 4481 4482 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P); 4483 4484 Changed = true; 4485 } 4486 4487 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(), 4488 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) { 4489 GenerateIVChain(*ChainI, Rewriter, DeadInsts); 4490 Changed = true; 4491 } 4492 // Clean up after ourselves. This must be done before deleting any 4493 // instructions. 4494 Rewriter.clear(); 4495 4496 Changed |= DeleteTriviallyDeadInstructions(DeadInsts); 4497} 4498 4499LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P) 4500 : IU(P->getAnalysis<IVUsers>()), 4501 SE(P->getAnalysis<ScalarEvolution>()), 4502 DT(P->getAnalysis<DominatorTree>()), 4503 LI(P->getAnalysis<LoopInfo>()), 4504 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) { 4505 4506 // If LoopSimplify form is not available, stay out of trouble. 4507 if (!L->isLoopSimplifyForm()) 4508 return; 4509 4510 // If there's no interesting work to be done, bail early. 4511 if (IU.empty()) return; 4512 4513#ifndef NDEBUG 4514 // All dominating loops must have preheaders, or SCEVExpander may not be able 4515 // to materialize an AddRecExpr whose Start is an outer AddRecExpr. 4516 // 4517 // IVUsers analysis should only create users that are dominated by simple loop 4518 // headers. Since this loop should dominate all of its users, its user list 4519 // should be empty if this loop itself is not within a simple loop nest. 4520 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader()); 4521 Rung; Rung = Rung->getIDom()) { 4522 BasicBlock *BB = Rung->getBlock(); 4523 const Loop *DomLoop = LI.getLoopFor(BB); 4524 if (DomLoop && DomLoop->getHeader() == BB) { 4525 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest"); 4526 } 4527 } 4528#endif // DEBUG 4529 4530 DEBUG(dbgs() << "\nLSR on loop "; 4531 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false); 4532 dbgs() << ":\n"); 4533 4534 // First, perform some low-level loop optimizations. 4535 OptimizeShadowIV(); 4536 OptimizeLoopTermCond(); 4537 4538 // If loop preparation eliminates all interesting IV users, bail. 4539 if (IU.empty()) return; 4540 4541 // Skip nested loops until we can model them better with formulae. 4542 if (!L->empty()) { 4543 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n"); 4544 return; 4545 } 4546 4547 // Start collecting data and preparing for the solver. 4548 CollectChains(); 4549 CollectInterestingTypesAndFactors(); 4550 CollectFixupsAndInitialFormulae(); 4551 CollectLoopInvariantFixupsAndFormulae(); 4552 4553 assert(!Uses.empty() && "IVUsers reported at least one use"); 4554 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n"; 4555 print_uses(dbgs())); 4556 4557 // Now use the reuse data to generate a bunch of interesting ways 4558 // to formulate the values needed for the uses. 4559 GenerateAllReuseFormulae(); 4560 4561 FilterOutUndesirableDedicatedRegisters(); 4562 NarrowSearchSpaceUsingHeuristics(); 4563 4564 SmallVector<const Formula *, 8> Solution; 4565 Solve(Solution); 4566 4567 // Release memory that is no longer needed. 4568 Factors.clear(); 4569 Types.clear(); 4570 RegUses.clear(); 4571 4572 if (Solution.empty()) 4573 return; 4574 4575#ifndef NDEBUG 4576 // Formulae should be legal. 4577 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), 4578 E = Uses.end(); I != E; ++I) { 4579 const LSRUse &LU = *I; 4580 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(), 4581 JE = LU.Formulae.end(); J != JE; ++J) 4582 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset, 4583 LU.Kind, LU.AccessTy, TLI) && 4584 "Illegal formula generated!"); 4585 }; 4586#endif 4587 4588 // Now that we've decided what we want, make it so. 4589 ImplementSolution(Solution, P); 4590} 4591 4592void LSRInstance::print_factors_and_types(raw_ostream &OS) const { 4593 if (Factors.empty() && Types.empty()) return; 4594 4595 OS << "LSR has identified the following interesting factors and types: "; 4596 bool First = true; 4597 4598 for (SmallSetVector<int64_t, 8>::const_iterator 4599 I = Factors.begin(), E = Factors.end(); I != E; ++I) { 4600 if (!First) OS << ", "; 4601 First = false; 4602 OS << '*' << *I; 4603 } 4604 4605 for (SmallSetVector<Type *, 4>::const_iterator 4606 I = Types.begin(), E = Types.end(); I != E; ++I) { 4607 if (!First) OS << ", "; 4608 First = false; 4609 OS << '(' << **I << ')'; 4610 } 4611 OS << '\n'; 4612} 4613 4614void LSRInstance::print_fixups(raw_ostream &OS) const { 4615 OS << "LSR is examining the following fixup sites:\n"; 4616 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(), 4617 E = Fixups.end(); I != E; ++I) { 4618 dbgs() << " "; 4619 I->print(OS); 4620 OS << '\n'; 4621 } 4622} 4623 4624void LSRInstance::print_uses(raw_ostream &OS) const { 4625 OS << "LSR is examining the following uses:\n"; 4626 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), 4627 E = Uses.end(); I != E; ++I) { 4628 const LSRUse &LU = *I; 4629 dbgs() << " "; 4630 LU.print(OS); 4631 OS << '\n'; 4632 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(), 4633 JE = LU.Formulae.end(); J != JE; ++J) { 4634 OS << " "; 4635 J->print(OS); 4636 OS << '\n'; 4637 } 4638 } 4639} 4640 4641void LSRInstance::print(raw_ostream &OS) const { 4642 print_factors_and_types(OS); 4643 print_fixups(OS); 4644 print_uses(OS); 4645} 4646 4647void LSRInstance::dump() const { 4648 print(errs()); errs() << '\n'; 4649} 4650 4651namespace { 4652 4653class LoopStrengthReduce : public LoopPass { 4654 /// TLI - Keep a pointer of a TargetLowering to consult for determining 4655 /// transformation profitability. 4656 const TargetLowering *const TLI; 4657 4658public: 4659 static char ID; // Pass ID, replacement for typeid 4660 explicit LoopStrengthReduce(const TargetLowering *tli = 0); 4661 4662private: 4663 bool runOnLoop(Loop *L, LPPassManager &LPM); 4664 void getAnalysisUsage(AnalysisUsage &AU) const; 4665}; 4666 4667} 4668 4669char LoopStrengthReduce::ID = 0; 4670INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce", 4671 "Loop Strength Reduction", false, false) 4672INITIALIZE_PASS_DEPENDENCY(DominatorTree) 4673INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 4674INITIALIZE_PASS_DEPENDENCY(IVUsers) 4675INITIALIZE_PASS_DEPENDENCY(LoopInfo) 4676INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 4677INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce", 4678 "Loop Strength Reduction", false, false) 4679 4680 4681Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) { 4682 return new LoopStrengthReduce(TLI); 4683} 4684 4685LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli) 4686 : LoopPass(ID), TLI(tli) { 4687 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry()); 4688 } 4689 4690void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const { 4691 // We split critical edges, so we change the CFG. However, we do update 4692 // many analyses if they are around. 4693 AU.addPreservedID(LoopSimplifyID); 4694 4695 AU.addRequired<LoopInfo>(); 4696 AU.addPreserved<LoopInfo>(); 4697 AU.addRequiredID(LoopSimplifyID); 4698 AU.addRequired<DominatorTree>(); 4699 AU.addPreserved<DominatorTree>(); 4700 AU.addRequired<ScalarEvolution>(); 4701 AU.addPreserved<ScalarEvolution>(); 4702 // Requiring LoopSimplify a second time here prevents IVUsers from running 4703 // twice, since LoopSimplify was invalidated by running ScalarEvolution. 4704 AU.addRequiredID(LoopSimplifyID); 4705 AU.addRequired<IVUsers>(); 4706 AU.addPreserved<IVUsers>(); 4707} 4708 4709bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) { 4710 bool Changed = false; 4711 4712 // Run the main LSR transformation. 4713 Changed |= LSRInstance(TLI, L, this).getChanged(); 4714 4715 // Remove any extra phis created by processing inner loops. 4716 Changed |= DeleteDeadPHIs(L->getHeader()); 4717 if (EnablePhiElim) { 4718 SmallVector<WeakVH, 16> DeadInsts; 4719 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr"); 4720#ifndef NDEBUG 4721 Rewriter.setDebugType(DEBUG_TYPE); 4722#endif 4723 unsigned numFolded = Rewriter. 4724 replaceCongruentIVs(L, &getAnalysis<DominatorTree>(), DeadInsts, TLI); 4725 if (numFolded) { 4726 Changed = true; 4727 DeleteTriviallyDeadInstructions(DeadInsts); 4728 DeleteDeadPHIs(L->getHeader()); 4729 } 4730 } 4731 return Changed; 4732} 4733